denoize
The pursuit of the world's highest-fidelity audio denoising — in pure Rust.
denoize removes background noise from WAV recordings with maximum transparency:
preserving timbre, transients, dynamics, stereo imaging, and natural "air".
Implemented technology stack
Classical DSP (always available)
- STFT/ISTFT + Perfect Reconstruction OLA + high overlap
- IMCRA/MCRA noise estimation + SPP + spectral-flatness profiling
- Ephraim-Malah Decision-Directed SNR
- 8 gain estimators: OMLSA, LogMMSE, MMSE-STSA, Wiener, SpecSub, SpecSub-NL, SpecSub-Geo
- Transient protection, cepstral smoothing, pre-emphasis
- Advanced windows: Kaiser, Flat-top, principal DPSS/Slepian (+ Hann/Hamming/Sine/Blackman)
- Multiband spectral subtraction (Bark bands)
- Perceptual weighting (Bark-scale gain shaping)
- Musical-noise post-filter
Optional AI backends (feature-gated)
| Backend | Feature | Description |
|---|---|---|
rnnoise |
--features rnnoise |
RNNoise via nnnoiseless (pure-Rust) |
deepfilter |
--features deepfilter |
DeepFilterNet v3 (tract ONNX, embedded model) |
onnx |
--features onnx |
External waveform-to-waveform ONNX model (tract, Pure Rust) |
mpsenet |
--features mpsenet |
MP-SENet magnitude/phase enhancement adapter (external converted model) |
bsrnn |
--features bsrnn |
ESPnet BSRNN spectral enhancement adapter (external converted model) |
mossformer2 |
--features mossformer2 |
ClearerVoice MossFormer2 48 kHz mask adapter (external converted model) |
sgmse |
--features sgmse |
SGMSE+ iterative diffusion adapter (external converted model) |
gtcrn |
--features gtcrn |
Official 48K-parameter causal GTCRN; reusable offline, --stream, library, and realtime sessions |
Build everything: cargo build --release --features full
The generic ONNX backend is the deployment foundation for future neural
models. Raw .onnx files and signed package v1 intentionally accept only
single-input/single-output waveform models. Signed package v2 additionally
authenticates named multi-input/output tensors, recurrent state, channel roles
and microphone geometry, latency/context, per-accelerator precision profiles,
resource budgets, provenance, and numerical conformance vectors. The generic
adapter executes only a finite-capable, independent-mono graph with one required
waveform input and one waveform output; expressive v2 graphs fail closed until
their dedicated restoration, target-speaker, AEC, or spatial adapter is
selected.
Library embedders can load OnnxWaveformModel
once, inspect its validated float32 waveform contract, and reuse the most
recently compiled input length instead of parsing and optimizing the graph on
every call. The CLI convenience path keeps the same --onnx-model contract.
The prebuilt GitHub binaries include every backend. Because the DeepFilterNet Rust crate is not available from crates.io, the crates.io package's
fullfeature currently includes RNNoise, generic ONNX, MP-SENet, BSRNN, MossFormer2, SGMSE+, GTCRN, and live-device support, but not DeepFilterNet. Building live support from source on Linux requires the ALSA development package (for example,libasound2-devon Debian/Ubuntu); prebuilt archives do not require compiler headers.
Hardware acceleration
CPU inference remains the compatibility default. Full builds register Apple Metal on Apple targets and NVIDIA CUDA on Linux/Windows targets for the generic ONNX, MP-SENet, BSRNN, MossFormer2, SGMSE+, and GTCRN adapters. Inspect the current binary and host without opening a model or using the network:
--accelerator auto uses the stable Metal-then-CUDA preference and reports an
explicit CPU fallback when the backend is CPU-only, deterministic mode is
active, or no compiled runtime passes its dependency probe. gpu, metal, and
cuda are strict requests. With an explicit backend, an unsupported or
unavailable request fails before input decoding; automatic backend selection
must inspect the decoded input before it can validate backend compatibility.
--deterministic always executes on CPU; combine it with cpu or auto.
File/stream JSON results expose requested, effective, and fallback, and
the effective runtime is included in batch recipe identity. The capability
contract is published as
denoize-hardware-v1.
For an available GPU the report also includes the device name, CUDA compute
capability when applicable, and the runtime-reported device-memory limit
(total global memory for CUDA; recommended maximum working set for Metal).
CUDA availability requires a compatible NVIDIA driver plus the CUDA runtime,
NVRTC, cuBLAS, cuDNN, CUDA development headers, CCCL headers, and a writable
tract kernel-cache directory. denoize hardware reports the first missing
prerequisite. The first CUDA model preparation can compile and cache kernels;
the host probe does not claim that every user-supplied ONNX graph is supported
by a GPU transform, so model preparation errors remain explicit.
Offline input and device recommendation
denoize recommend analyzes the signal, compiled backends, locally installed
models authenticated by the embedded signed catalog, resource limit, and
current CPU/GPU availability without updating the catalog or model cache,
downloading a model, or contacting a service:
The command analyzes at most 12 seconds by default (configurable from 1 to 60).
WAV, FLAC, and Ogg Vorbis are analyzed through the bounded block decoder; other
supported formats use their existing whole-file decoder under the same
--max-memory ceiling and then analyze only the bounded prefix. Filesystem
inputs retain the regular-file and same-opened-object guarantees. Recommendation
still requires a filesystem input; - is reserved for bounded --stream
processing rather than recommendation.
--calibrate adds a fixed, hash-identified Classical Hi-Fi workload with one
warmup and three measured runs after its fixed scratch allowance fits the same
memory ceiling. The report records raw timings and median
baseline realtime headroom, then combines that evidence with documented
backend cost classes; it does not pretend to be a full-reference audio-quality
benchmark or an exact runtime prediction for a neural model. Every candidate
includes stable reason codes, eligibility, local model/runtime state, estimated
denoize-owned CPU and GPU memory, and the suggested explicit arguments. GPU
eligibility respects both --max-gpu-memory and a runtime-reported device limit
when one is available. Recommendation captures one read-only hardware snapshot
and does not create or test a CUDA kernel cache; actual processing revalidates
cache writability before preparing a model. Compact and pretty output use the
versioned
denoize-recommendation-v1
contract. Backends that require a caller-supplied model path remain visible as
excluded candidates; the report does not disclose paths, so only configurations
that can be reproduced without inventing a model argument are auto-selected.
The analyzed-sample SHA-256 is a content fingerprint; remove it before sharing
a report when correlating the source audio would be sensitive.
Native degradation diagnosis and no-reference assessment
denoize diagnose performs bounded, deterministic, network-free triage when a clean reference is unavailable:
The report separates additive noise, clipping, 50/60 Hz hum, clicks, reverberation, bandwidth limitation, short dropouts, wind/plosive energy, and codec risk. Each finding carries direct evidence, continuous severity, confidence, and a recommended restoration action. assess can evaluate one input or compare before/after quality while separately checking sample rate, channel count, and presentation duration.
The native score and MOS proxy are triage estimates, not human MOS. Schema v1 always reports semantic fidelity as unassessed; it cannot prove preservation of words, phonemes, speaker identity, language, or prosody and cannot authorize a generative result by score alone. See Native diagnostics and the closed denoize-diagnostic-v1 and denoize-assessment-v1 contracts.
Deterministic audio restoration
denoize restore provides conservative, model-free de-clipping, prediction-residual de-clicking, harmonic de-hum, finite WPE de-reverberation, and short wind/plosive repair:
The pipeline never changes decoded sample rate, channels, or frame count. Detect-only mode is bit-exact and can export a complete channel/frame RLE mask without writing audio. Apply mode uses conservative confidence gates, explicit iteration and attenuation ceilings, bounded memory admission, atomic no-clobber outputs, and a canonical operation order. Reports contain PCM and mask digests, detected/changed counts, confidence, energy delta, warnings, and closed per-operation evidence, but no filesystem path. Context padding is separate from accepted damage and from samples actually replaced.
See Deterministic restoration, the closed report and mask contracts, and the research and acceptance review.
Fail-closed universal speech restoration
denoize universal runs a signed runtime package v2 through the dedicated
48 kHz BSRNN spectral adapter. The discriminative primary path is the safe
default; hybrid and generative packages require both an alternate role and an
explicit experimental opt-in:
The package signature, provenance, exact graph interface, selected-runtime resource profile, and numerical vectors pass before source inference. Clean input bypasses the model. A private candidate is published only when geometry, finite-sample, energy, peak, new-clipping, silence-injection, and native-quality gates all pass; otherwise the decoded input is written unchanged. Reports bind package/key/source/checkpoint and PCM/mask SHA-256 values without paths.
Signal gates do not prove preservation of words, phonemes, prosody, or speaker identity. Model promotion therefore uses separately signed evidence covering 20 required demographic/material/degradation strata, nine metrics per stratum, and human-listening thresholds. Upstream URGENT and UniPASE weights are not bundled because their complete artifact-level training-data redistribution chain has not been established. See Universal restoration, its research audit, and the closed report, mask, and promotion evidence contracts.
Fail-closed target-speaker extraction
denoize target-speaker extracts one enrolled speaker from a mixture through a dedicated signed package v2 graph. Stage 29 provides a fail-closed offline renderer and a separately gated causal mono renderer:
The graph must expose exactly one mixture input, one enrollment input, one
same-length audio output, and calibrated absent/uncertain/present
probabilities. Package components, provenance, graph names/shapes, numerical
vectors, and separately signed REAL-T/TS-SUPERB/absence promotion evidence are
verified before either audio file is decoded.
Only a confidently present target whose candidate passes geometry, finite, energy, peak, clipping, presence, and evidence gates creates an audio file. Absent, uncertain, and unsafe candidates create no audio and never fall back to the mixture, silence, or an unverified voice. Enrollment working buffers are zeroized immediately after inference; reports contain no enrollment samples, embedding, digest, or path.
The causal command additionally requires accepted offline evidence and a signed 22-stratum non-inferiority, recurrent reset/flush, <=100 ms perturbation latency, 10,000-block callback, and target-transition audit:
It authenticates both evidence layers and recurrent vectors before decoding, removes only the signed latency after a complete flush, preserves exact source duration, and renders silence for absent, uncertain, warm-up, or unsafe blocks. The public real-time bridge uses fixed lock-free queues, discards late/stale results, and keeps inference off the callback; no plug-in consumes enrollment until its own privacy and real-host gate passes.
No checkpoint is bundled because the audited WeSep/REAL-TSE/MeanFlow candidates do not yet provide a complete artifact-level redistribution and protected- stratum evidence chain. See Target-speaker extraction, the paper and artifact audit, and the closed report and promotion evidence contracts, plus the causal report and promotion evidence contracts.
Fail-closed acoustic echo cancellation
denoize aec cancels one typed mono far-end playback reference from one
microphone recording through the native aec feature:
The exact closed configuration and native implementation are bound by separately signed promotion evidence before either audio file is opened. The safe path uses explicit constant-clock mapping, normalized FFT signed-delay estimation, partitioned frequency-domain NLMS, double-talk adaptation freeze, and conservative residual suppression. A missing or low-confidence reference preserves the microphone. Route, reference, clock, delay, and non-finite-state boundaries cold-reset the filter; output retains the exact microphone rate and frame count.
AecStream plans FFTs and preallocates every filter/history/scratch block on
the control thread. AecRealtimeAdapter adds arbitrary host quanta with one
fixed AEC-block latency and preallocated typed microphone/reference buffers.
Processing allocates, locks, waits, logs, and performs I/O zero times. Promotion
requires the complete 17-stratum
delay/drift/path/talk/impairment matrix, <=20 ms latency, worst-case RTF <=0.5,
10,000 paced blocks, real-device and nonlinear cases, AECMOS/WAcc/listening,
bounded reconvergence, and zero callback/deadline/stale-reset violations. ERLE
is reported only for far-end-only blocks. See
Acoustic echo cancellation, its
paper and artifact audit,
and the closed report and
promotion evidence
contracts.
Explicit-geometry microphone-array enhancement
denoize array enhances a declared two-to-four-channel microphone array to mono without ever inferring that ordinary program stereo is an array:
The closed configuration binds channel IDs, unique right-handed coordinates, sample-skew and gain/phase calibration, the reference microphone, and every DSP parameter. Signed promotion evidence authenticates that exact configuration before audio is opened. The native baseline applies multichannel WPE followed by conditioned mask-MVDR; singular bins and arrays with too few active channels use the declared reference, while an inactive reference fails closed. Channel plus geometry permutations render equivalently, output is finite mono PCM of exact duration, and reports record no paths.
No neural spatial checkpoint is bundled or promoted; streaming and moving- source neural claims remain behind the package-v2 and real-device evidence gates. See Microphone-array enhancement, the paper and artifact audit, and the closed report and promotion evidence contracts.
Supported input formats
| Format | Decoder | Notes |
|---|---|---|
| WAV/BWF | hound |
8–32 bit int / float; BWF metadata chunks are preserved for supported tags |
| RF64 | native RF64 reader | 64-bit-size PCM/WAVE, bounded chunk reads |
| AIFF/AIFC | symphonia |
PCM and supported AIFC codecs |
| CAF | symphonia |
PCM and ALAC/other supported CAF codecs |
| MP3 | symphonia + bounded nanomp3 fallback (Pure Rust) |
Xing/Info + LAME gapless trim, ID3v2, no resampling |
| M4A/AAC/ALAC | oxideav-aac + symphonia fallback |
AAC-LC/ALAC decode with MP4 v0/v1 unity-rate edit-list timing; MP4 AAC-LC uses the 24-bit MPEG-4 buffer/sample safety ceiling and charges payload-proportional decoder work before access-unit allocation when --max-memory is set (ALAC is unaffected) |
| FLAC | claxon |
Lossless FLAC |
| Ogg Opus/Vorbis | opus + ogg / symphonia |
Mono/stereo; native sample rate decode |
Output formats
| Format | Encoder | Notes |
|---|---|---|
| WAV | hound |
Lossless; preserves bit depth |
| MP3 | shine-rs (Pure Rust) |
--mp3-bitrate (default 192 kbps) |
| M4A | oxideav-aac + MP4 mux |
GitHub/source builds; positive --m4a-bitrate (default 192 kbps) |
| FLAC | flacenc |
Lossless, pure Rust |
| Ogg Opus | opus + ogg |
128 kbps, mono/stereo |
MP3 inputs with a Xing/Info header and a compatible LAME, Lavf, or Lavc extension are decoded over their exact signalled presentation span: encoder delay and end padding are not exposed to the denoising pipeline. Untagged raw MP3 has only MPEG-frame timing, so its decoded duration remains frame-rounded. Symphonia remains the primary decoder; a fixed-size streaming compatibility fallback is used only for its exact invalid-bit-reservoir error on untagged, contiguous Layer III frames, and never for files carrying gapless timing. The built-in Shine encoder completes its final bit cache and emits at least two MPEG frames for short-clip interoperability; clips shorter than that minimum therefore contain trailing encoded silence.
M4A AAC-LC and ALAC inputs are rendered on the container presentation timeline. Unity-rate v0/v1 edit lists may trim or splice media and insert leading or interior silence; unsupported rates and malformed timing are rejected instead of returning mis-timed PCM.
Channel order is kept planar and unchanged through WAV/FLAC and denoising. The
standard layouts mono, stereo, 2.1, quad, 5.0, 5.1, 6.1, and 7.1 are reported
when their channel count is recognized. MP3, M4A, and ADTS AAC encoders in the
current release accept only mono/stereo; surround input is rejected instead of
being mixed implicitly. Use --downmix stereo when a documented, explicit
surround-to-stereo render is intended (LFE is not copied into the full-range
stereo pair). WAVE_FORMAT_EXTENSIBLE speaker masks are read, preserved, and
written for multichannel WAV files; --report also shows each channel's
azimuth/elevation pan coordinate. A non-standard but valid mask is used for
position-aware downmixing instead of being guessed from the channel count.
Stereo processing can be selected with --channels mid-side. This uses a
reversible, energy-preserving Mid/Side transform (M=(L+R)/sqrt(2),
S=(L-R)/sqrt(2)) and reconstructs the original channel order and speaker
metadata after denoising.
# MP3 / M4A input and output — no manual ffmpeg conversion
# User-supplied waveform model: [1, samples] or [1, 1, samples]
# Signed model + license + frontend/tensor/resource contracts
# Inspect either signed package version; v2 also reports named I/O, state,
# latency, precision profiles, provenance, and numerical-vector coverage
# Official MP-SENet checkpoint converted with scripts/export-mpsenet.py
# ESPnet BSRNN xtiny checkpoint converted with scripts/export-bsrnn.py
# ClearerVoice MossFormer2 48 kHz model
# Official SGMSE+ VoiceBank model (30-step quality sampler)
# Verified official GTCRN model (manual model path is unnecessary afterwards)
# Stereo coupling, pipes, metrics, and directory batches
|
To prepare the pinned official MP-SENet VoiceBank model:
The VoiceBank graph is about 9 MiB and expects 16 kHz audio. On the reference x86-64 Linux host, a two-second mono speech fixture took 43.67 seconds after model loading and the complete process used 410,048 KiB maximum RSS. Run the pinned real-speech quality gate after conversion:
To prepare the pinned ESPnet BSRNN xtiny model (CC-BY-4.0):
|
The adapter resamples to 48 kHz and reproduces the published model's variance normalization, centered 960-point Hann STFT with a 480-sample hop, whole-utterance recurrent inference, and inverse STFT. The converted model is about 2.4 MiB. On a release build on the project reference x86-64 Linux host, the fixed two-second regression fixture took 1.58 seconds (1.3x realtime) and used 44,628 KiB maximum RSS. Runtime and memory grow with utterance length.
Run the reproducible real-speech quality gate after conversion:
To prepare the pinned Apache-2.0 MossFormer2 SE 48 kHz model:
|
The adapter uses 48 kHz audio, 40 ms Kaldi fbank frames at an 8 ms shift, first- and second-order deltas, a non-centred 1,920-point symmetric-Hamming STFT, and the official four-second/three-second-stride edge-discard reconstruction. The converted graph is about 217 MiB. On the reference x86-64 Linux host, a four-second mono fixture took 7.74 seconds and used 483,400 KiB maximum RSS in a release build. Model weights are not bundled.
Run the pinned real-speech quality gate after conversion:
To prepare the pinned MIT-licensed SGMSE+ VoiceBank+DEMAND model:
|
The adapter reproduces the official noisy-peak normalization, centered
510-point periodic-Hann STFT with a 128-sample hop, complex square-root
spectrum transform, and OUVE predictor/corrector sampler. The explicit
quality/speed choice is the upstream 30 reverse steps with one ALD corrector
step (snr=0.5), or 60 score-network evaluations. The graph is about 252 MiB
and weights are not bundled. On the reference x86-64 Linux host, the pinned
two-second mono fixture took 737.92 seconds and used 1,204,648 KiB maximum RSS
in a release build. This backend prioritizes generative quality rather than
interactive speed.
Run the pinned quality gate after conversion (expect a long CPU run):
Quick start
# Best classical quality
# RNNoise AI backend
# DeepFilterNet v3 AI backend
# Advanced DSP options
# Principal periodic DPSS/Slepian taper (NW must be in (0, 8])
For DPSS, NW is the time-bandwidth product and defaults to 3.0. Increasing
it widens the main lobe and concentrated frequency band while strengthening
the taper toward the frame edges. DPSS is a classical-backend STFT option; the
equivalent TOML is:
= "classical"
= "dpss"
= 3.0
Managed model downloads
Every build contains a versioned model catalog and trust-root policy. Remote
catalog updates are accepted only after detached-minisign verification against
the active root, strict schema and expiry validation, and monotonic
sequence/rollback checks. Trust-root rotations advance exactly one version and
must satisfy distinct-signature thresholds from both the current and candidate
root; catalog signing keys can be given closed sequence windows or explicit
revocation cutoffs, while rotations cannot weaken the active expiration policy.
Model
installs and updates then use the active catalog's exact size and SHA-256 while
supporting explicit network policy, authenticated mirrors, resumable transfers,
and air-gapped local files. Run denoize models --help for the dedicated
command reference.
# Inspect, update, or air-gap import the signed catalog.
# Inspect, rotate, or recover catalog trust. Rotations use a JSON bundle of
# detached signatures from the current and candidate root key sets.
# Only after correcting an accidental future system-clock jump:
# Verify and import one release bundle without opening a network connection.
# Diagnose the whole cache without changing model data. Repair known packages,
# then preview and apply removal of stale denoize-owned state.
# Offline mode never opens a network connection.
# Override the source and proxy for one model.
# Read origin credentials from the environment, never a CLI secret value.
# Ignore all proxy settings, or install an already transferred local file.
The corresponding defaults are DENOIZE_MODEL_OFFLINE,
DENOIZE_MODEL_URL, DENOIZE_MODEL_CATALOG_URL, DENOIZE_MODEL_PROXY,
DENOIZE_MODEL_BEARER_TOKEN, DENOIZE_MODEL_USERNAME, and
DENOIZE_MODEL_PASSWORD. Standard HTTPS_PROXY, HTTP_PROXY, ALL_PROXY,
and NO_PROXY variables (including lowercase variants) are used when no
denoize-specific proxy override is active. --proxy selects an explicit
proxy; --no-proxy and an empty DENOIZE_MODEL_PROXY force a direct
connection.
Interrupted transfers are retained in a .part sidecar and resumed with HTTP
range requests. Saved ETag or Last-Modified validators and each
Content-Range are checked before appending; changed objects, malformed range
responses, or an unverified 416 response cause a clean restart. Every managed
model candidate must match both the catalog's exact byte length and SHA-256
before use: this includes fresh or resumed downloads, alternate --url
sources, --from imports, completed partials, and files already in the cache.
An update keeps the current verified model until its replacement is ready.
Each installation also receives content-addressed provenance that binds the
artifact to its catalog sequence, digest, and signing key, and records its
installation-time catalog origin and source. Existing verified caches are
migrated lazily; mismatched provenance fails verification. denoize models info MODEL reports these fields and the
pinned length as an unscaled decimal size-bytes value. This per-model
integrity bound is not an aggregate cache quota.
Official releases additionally provide a signed offline .dmb bundle for
closed networks. Its bounded, length-delimited format carries the exact catalog,
detached signature, trust root, every catalog model, upstream license text, and
source-provenance JSON. The manifest contains no extraction paths and the format
has no compression layer. models bundle inspect authenticates every byte
without changing catalog or cache state; models bundle import performs the
same full preflight before activating the catalog and atomically installing any
missing model. Both commands accept only a regular file and perform no network
I/O. A model installed this way records the bundle SHA-256 in local installation
provenance.
Catalog expiry and rollback rules still apply offline. The import may advance
the monotonic catalog floor before a later storage error; it rolls back model
artifacts created by that invocation, and retrying the same authenticated bundle
is safe. Existing valid models are retained. See the
managed-model guide for the operator
layout used by models bundle create and the complete transaction contract.
Catalog sequence 1 is a compatibility exception because it predates signed
timestamps. The embedded v1 trust policy requires every later sequence to carry
issued_at_unix_seconds and expires_at_unix_seconds, with at most 180 days of
validity. The root itself expires at its displayed Unix time. denoize persists
the greatest trusted wall-clock value it has observed, so setting the system
clock backwards cannot reactivate expired authority. Expiry, a newly tightened
timestamp policy or key window, or revoked_at_sequence stops new installs,
updates, local imports, and artifact reacquisition; already installed bytes
remain usable and verifiable, and local provenance-only repair, diagnosis,
pruning, and removal remain available.
models catalog trust recover replaces corrupt or incomplete cached trust
metadata only with the root compiled into the running binary. It never lowers a
valid newer root or the catalog rollback floor; those cases require the missing
signed chain or a newer denoize binary. A newer embedded root is therefore the
independent emergency recovery channel. By default recovery also preserves the
greatest trusted time already observed. After correcting an accidental future
system-clock jump, the explicit models catalog trust reset-time-floor command
resets only that clock floor to the current system time while retaining the
active signed root and chain. It cannot lower either the trust-root version or
catalog rollback floor, and refuses the reset unless the active root is valid at
the corrected current time. Inspect the recorded value with models catalog trust status before using the command.
models doctor inventories every active-catalog package plus cache sidecars
and orphan entries without changing artifacts, provenance, or download state.
An optional package that was never installed is reported as missing but does
not make a fresh cache unhealthy. Corrupt bytes, missing or mismatched
provenance, incomplete/stale downloads, unsafe entries, and catalog-orphaned
packages are reported separately. models verify MODEL|all remains the strict
package verification command.
models repair MODEL|all rebuilds provenance locally when verified bytes are
already present; otherwise it uses the same offline, source, proxy, and
authentication policy as install/update. Replacement bytes are staged and
verified before the old artifact is atomically replaced. models prune --dry-run lists exact removable paths. Applying models prune deletes stale
sidecars, superseded provenance, and old package directories only when their
content-addressed provenance, artifact digest/size, and directory layout all
match denoize-managed state. Unknown data and symlinks, devices, or other
special entries are reported and retained.
models snapshot [--json] [--pretty] emits the stable
denoize-automation-v1 document without network access. It binds the active
catalog and trust root to cache health, every expected model identity, validated
installation provenance, and the processing recipe ABI in one
generation-checked snapshot. Normal --json processing and batch NDJSON records use
denoize-cli-output-v1 and expose each finite-file recipe digest. The desktop
model library exports the same automation snapshot atomically. See the
stable JSON contract and its versioned schemas for field and
compatibility rules.
HTTPS model connections, including those tunneled through an HTTP CONNECT
proxy, use the operating system trust store. CLI Bearer tokens and Basic
passwords are accepted through environment variables, and diagnostics redact
credentials, query strings, and fragments. Signed --url values and proxy
credentials can still leak through process listings and shell history, so use
protected environment injection when that matters. See the
managed-model guide for option combinations,
proxy precedence, and resume validation details.
Long recordings with bounded memory
For long recordings, use the stateful streaming path. It keeps only a fixed-size input block plus bounded decoder, backend overlap, recurrent, resampler, and encoder state in memory instead of loading the whole file:
--stream accepts content-detected WAV, FLAC, Ogg Vorbis, granule-aware Ogg
Opus, gapless MP3, frame-aware ADTS AAC, and edit-aware M4A AAC/ALAC input. It
writes WAV, FLAC, Ogg Opus, MP3, M4A AAC, or ADTS AAC with the compiled
Classical, RNNoise, DeepFilterNet, MossFormer2, and GTCRN backends, including
independent, linked-stereo, and mid/side channel modes. MossFormer2 and GTCRN
use an explicit --onnx-model or their installed managed model; DeepFilterNet
uses its embedded graph. A regular-file destination is staged, decoded
end-to-end to verify its codec, geometry, and presentation duration, then
published atomically. Supported metadata is retained unless --no-metadata is
selected. Bounded VAD preserves the presentation timeline across backend
latency. --loudness performs a fixed-memory analysis pass over an anonymous
PCM spool, then applies one constant gain during the verified encoding pass.
The default block size is 8192 frames; use --stream-frames N (1–1,048,576) to
trade latency and working memory for throughput. Noise profiling retains only
a bounded leading segment before output begins. Stream resource arithmetic is
checked from the input header, and the processor is constructed before an
output or temporary file is staged.
- selects stdin or stdout for --stream. Stdin is copied into an anonymous
regular-file spool before parsing, preserving one authoritative seekable input
without retaining all encoded bytes in RAM. Stdout accumulates bounded PCM and
encoded anonymous spools, applies metadata and optional two-pass loudness,
verifies the completed encoded stream, then copies it to the sink. Stdin and
stdout share the --max-temp-space allowance (1 GiB by default); supported
input metadata is preserved unless --no-metadata is selected. A sink error
can leave partial external bytes because a pipe has no atomic rename. --resume
intentionally rejects stdin/stdout because their anonymous spools cannot
survive process restart.
Library callers use AudioStreamReader::from_reader_with_limits for a plain
Read source and SpooledAudioStreamWriter::new_with_limits for a plain
Write sink. StreamSpoolLimits bounds encoded input bytes or, for output,
the simultaneous PCM spool, encoded spool, and codec auxiliary files. The
seekable AudioStreamWriter remains the lower-overhead choice when the caller
already owns a private Write + Seek transaction.
Add --resume to make a long stream restartable. The CLI periodically
synchronizes a private append-only checkpoint journal and an interleaved f64
PCM spool beside the destination. After interruption it deterministically
replays the same opened input to the last durable boundary, verifies the saved
PCM digest, restores backend state, and continues. The input bytes, effective
recipe, model bytes, source format, channel geometry, and block size are bound
to the checkpoint. A mismatch is preserved and rejected unless --force
explicitly discards it. Codec delay, Ogg granules, and M4A edit lists are
applied before presentation PCM reaches a durable boundary. The final encoded
output is staged, decoded for verification, and committed atomically; success
removes the journal and PCM spool while retaining the reusable lock file. The
journal records the exact verified staged-output fingerprint before
publication. If the process exits after the atomic commit but before receipt
publication or sidecar cleanup, the next identical resume verifies the
destination, reports a skip/completed plan when requested, can publish the
matching signed receipt, and removes the stale data sidecars without
reprocessing. A changed destination is preserved and rejected unless --force
resets the checkpoint. The PCM spool, staged encoded file, encoder auxiliary
data, and retained metadata are all charged to --max-temp-space.
Filesystem inputs are opened once per processing phase as validated regular files. Size estimation, probing, decoding, and metadata reads within that phase use the same opened filesystem object, so replacing the pathname cannot silently mix bytes from two inputs. FIFOs, directories, and device files are rejected before an audio parser or output staging step runs.
For the normal (decoded, non-streaming) path, --max-memory MB caps requested
denoize-owned decoded PCM capacities and explicitly accounted codec scratch
buffers, in addition to the conservative input-size preflight and final
decoded-working-set check. --max-process-memory MB adds weighted admission
across all active workers and retained model sessions. Batch preflight reserves
each decoder's complete configured allowance, rechecks it after model loading,
and starts a worker only when its RAM, temporary-output, CPU, and GPU request
fits atomically. Actual retained metadata is charged conservatively as well.
--max-temp-space MB bounds aggregate staged-output reservations (including a
restartable stream's PCM spool) and verifies the final staged length before
publication; it is not a filesystem quota.
--max-gpu-jobs N (default 1) serializes or bounds accelerated workers, while
--max-gpu-memory MB applies conservative model and transfer reservations.
Those GPU counters are not driver-reported VRAM usage. Internal allocations
made inside third-party codec or model runtimes can still fall outside the
cooperative counters, and allocator capacity rounding means they are not an
allocator-exact process RSS limit.
For an OS-enforced CLI boundary, add --isolate. Processing runs in a child;
on Unix, --max-process-memory becomes an RLIMIT_AS address-space ceiling,
and on Windows it becomes a Job Object process-memory ceiling. The parent
survives a decoder/model abort and publishes no staged output from a failed
child. Without a process-memory value, isolation still contains child failure
but does not invent a memory ceiling. Desktop file and batch jobs always use
the equivalent supervised child boundary; desktop live audio remains in
the application process under the cooperative governor.
FLAC and Ogg structure is also checked with finite block,
packet, page, stream, item, and aggregate metadata limits before a decoder can
materialize it—even with --no-metadata. When tags are preserved, their
retained payload budget is derived from the memory left after the decoded PCM
working set; the same limit is enforced again while writing the staged output.
The default limits remain finite when --max-memory is omitted.
The per-input limit applies to regular-file inputs/workers; stdin retains its
separate bounded WAV buffering path. Process admission makes --jobs an upper
bound rather than a promise that every worker can run simultaneously. Batch
probing, decode, model preparation, and metadata validation all finish before
the output directory or staging files are created. A streaming job stays
bounded by its block size, decoder allowance, and denoiser state, and metadata
uses a conservative share of the remaining budget:
DAW plug-ins
The denoize.clap bundle exposes two effects with independent stable IDs. The
fixed-memory DSP effect is org.penguin425.denoize. It supports mono and stereo
ports, f32 and f64
audio, in-place and out-of-place buffers, sample-accurate automation, bypass,
stereo linking, dry/wet mix, and output gain. Activation allocates all delay
and DSP state up front. The audio callback performs no allocation, locking,
filesystem or network I/O, or system calls.
The plug-in reports the fixed fixed-10ms-v1 latency policy to the host. The
exact frame count is ceil(sample_rate * 0.010): 441 frames at 44.1 kHz, 480
at 48 kHz, and 960 at 96 kHz. denoize plugin latency independently sends an
impulse through the bypassed f64 processor and fails if the measured first
output frame differs from the reported value:
The model-backed effect is denoize Neural
(org.penguin425.denoize.neural). Install the pinned managed gtcrn-dns3
graph before a DAW activates it; the host process never downloads a model:
Neural inference, graph preparation, resampling, and recurrent state run on one
permanent worker. The callback only copies through preallocated blocks and
bounded lock-free queues; it performs no model work, allocation, locks, waits,
I/O, network access, or logging. Its fixed policy is
fixed-24x10ms-worker-v1: 24 * ceil(sample_rate * 0.010), or 240 ms at
ordinary rates. Finite fractional CLAP sample rates are supported. A late,
invalid, or missing result uses latency-aligned dry audio by default; last-safe
gain and silence require an explicit parameter choice. The advertised
reference input is reserved for later typed target-speaker/AEC semantics.
Portable presets use
denoize-daw-preset-v1. Complete
session state uses
denoize-daw-session-v1 and
binds the plug-in ID, latency policy, mono/stereo port configuration, and every
parameter. Both formats reject unknown fields and future versions, are bounded
to 64 KiB, read only regular non-symlink files, and publish atomically with
no-clobber as the default. CLAP host state serializes the same deterministic
session document, so standalone files and DAW restoration cannot drift into
separate state formats.
Neural state uses
denoize-neural-daw-session-v1
and additionally binds the exact model ID, graph SHA-256, overload fallback,
and scheduler policy. It is also path-free, closed, bounded, non-symlink, and
atomically no-clobber. See Neural DAW plug-in for the
scheduler, model trust boundary, state contract, evidence, and limitations.
From a repository checkout, build a local CLAP plug-in with
cargo build --release -p denoize-clap. On Linux,
copy target/release/libdenoize_clap.so to ~/.clap/denoize.clap. Tagged
releases provide ready-to-copy archives for Linux x86-64, macOS Intel and Apple
Silicon, and Windows x86-64. macOS archives contain a complete
denoize.clap bundle. After copying it to a standard CLAP directory, restart
the DAW or run its plug-in rescan. CI and the tagged release workflow verify
both descriptors with the pinned official CLAP validator 0.4.1: 81 tests, 68
applicable passes, no failures or warnings, and 13 capability-based skips.
v0.79.0 adds one accessible native embedded editor for both CLAP descriptors. Every visible control remains a host parameter, with keyboard navigation, visible focus, deterministic software rendering, native AccessKit adapters, bounded lock-free UI automation, and complete generic-host fallback when the custom window API is unsupported or creation fails. The signed Linux X11 gate opens and renders both editors in a real host, injects a bypass click, verifies the exact three-event automation gesture, and exercises resize and lifecycle rejection paths. See Accessible plug-in editor for the supported window APIs, accessibility contract, evidence, and explicit limits.
VST3 3.8 bundles are available from v0.78.1. They statically adapt the same two descriptors through exact pinned CLAP-wrapper, CLAP SDK, and VST3 SDK revisions, so they cannot load a same-named external CLAP binary. The official 3.8.1 validator gate requires 94/94 passes. A pinned Ardour 8.4 headless real-host gate also discovers, inserts, processes, saves, reloads in a fresh process, and tears down both descriptors at 48 kHz. Release evidence includes a signed host matrix and both bound logs. See VST3 plug-in for installation, reproducible build details, and the explicitly unclaimed f64, custom-view, and proprietary-host cases. The v0.79.0 editor evidence is native CLAP-only and does not silently widen those VST3 claims.
v0.80.0 adds a macOS AUv3 app extension for both descriptors, with stable
aufx/Dn01/Dnze and aufx/Dn02/Dnze identities. The signed containing app
embeds the sandboxed extension, exact signed CLAP, and verified GTCRN graph plus
authenticated provenance, so Neural does not depend on a cache hidden by the
app-extension sandbox. Apple auval and an independent AVFoundation
lifecycle/state host exercise both components on Intel and Apple Silicon; the
target-qualified evidence and reports are signed. See
AUv3 plug-in for installation, reproducible builds, and
the explicit iOS, proprietary-host, and custom-view limits.
v0.81.0 adds direct Linux x86-64 LV2 adapters for both descriptors. Neural uses the host-owned LV2 Worker instead of a private inference thread; bounded Atom/Patch events provide sample-accurate automation, and the State interface round-trips the same closed portable JSON used by the shared engine. The release gate combines official LV2 metadata validation and Lilv discovery with real Jalv Worker processing and a two-process Ardour save/reload test. One signed evidence record binds the exact descriptor URIs, port counts, fixed latencies, package versions, source commit, and three raw reports. See LV2 plug-in for installation and the explicit Linux-only, f32-only, generic-editor, and untested-host limits.
Desktop app
The Tauri desktop app exposes single-file denoising, batch conversion, portable
project timelines, DAW plug-in state management, native degradation diagnosis,
no-reference before/after assessment, signed quality comparison, and model
management without sending audio off the computer. Its
default build includes every backend in the repository's full feature set;
FDK-AAC remains an explicit opt-in because of its separate licensing terms.
ONNX-based backends expose model-file, model-rate, and SGMSE quality controls
when selected; managed GTCRN weights are resolved automatically after install.
The resource panel applies aggregate RAM and staged-output admission, a
conservative GPU-memory reservation, and a GPU-worker concurrency limit to
single-file, batch, and live jobs. Final file, batch, and short preview work run
in supervised child processes so a decoder or model abort cannot directly take
down the UI. Unix workers disable core dumps and die with their parent; Windows
workers start behind a gate, enter a kill-on-close Job Object before processing,
and apply the configured process-memory ceiling. Without an explicit ceiling,
isolation still contains a worker crash but does not claim allocator-exact RSS.
Final workers exchange bounded nonce-authenticated progress records. A shared
commit/cancel fence prevents a cancelled or rejected worker from publishing a
later output.
The model manager shows signed-catalog identity and installed provenance and
can update the catalog or atomically export the stable automation JSON. Its
offline, alternate-source, proxy/direct,
authentication, and local-file controls are session-only. Bearer tokens and
Basic credentials are cleared after an operation starts, and none of these
download overrides are included in saved settings, named presets, or
CLI-compatible imports and exports.
Desktop batches accept files or folders, preserve relative paths, run with a
configurable worker count, continue after individual failures, and can resume
from the same .denoize-state journal used by the CLI in the output directory.
Single-file processing also provides a bounded non-destructive audition flow.
It renders at most 30 seconds and up to three candidate recipes, exposes
loudness-matched original, processed, and removed-signal audio, supports
keyboard seeking and looping, and includes a blind A/B choice. A selected
recipe is persisted locally, but applying it to a final job still requires the
same source fingerprint, effective backend, output format, and recipe. Restored
choices must be rendered again before use. The public
denoize-presentation-region-v1
locator stores exact presentation ticks rather than encoded packet time. A
cancelled or failed preview publishes no final output or restart state and its
private temporary directory is removed.
Desktop settings are restored automatically, can be stored as named presets,
and can be imported or exported as CLI-compatible TOML. Recent input files are
kept locally for quick reuse. The single-file and batch views also expose a
reproducibility mode that serializes processing and uses stable model seeds.
The single-file view can also run the bounded WAV, FLAC, Ogg Vorbis/Opus, MP3,
ADTS AAC, and M4A AAC/ALAC input path, choose any supported encoded output and
block size, preview the read-only v2 execution plan, publish a signed receipt,
and enable the same durable restart checkpoints as the CLI.
Audio files and folders can be dropped onto the single-file or batch input
zones; output folders have dedicated drop targets. Multiple audio files switch
the app to batch mode automatically.
The realtime page routes a selected capture device through a low-latency
backend to a playback device, with independent input/output sample rates,
adaptive clock correction, bounded playback priming, and explicit start/stop
controls. It reports queue depth, estimated capture-to-playback latency, drift
correction, underrun/overflow frames, dropped chunks, and reconnect attempts.
If a device disappears, the session retries the selected name (or the current
system default) for the configured finite recovery window. Live sessions
support the live-capable Classical, RNNoise, and GTCRN backends; other backends
are rejected before capture or playback starts. GTCRN requires its managed
model to be installed (or an explicit model path) and keeps one optimized graph
with independent recurrent state per processed channel. Headphones help
prevent acoustic feedback.
The DAW plug-in page displays the exact reported and measured latency for both DSP and Neural effects at a selected sample rate. It also shows the pinned neural model/install state, bounded queue and overload policy, loads the three DSP factory presets, edits every stable DSP parameter, and imports or atomically exports portable preset and deterministic session JSON. All validation and file publication remain in Rust; the WebView does not implement a second parser or state contract.
The Project page loads and validates source-bound manifests, selects a linear sample-accurate timeline, previews or saves its exact execution plan, assembles a verified float WAV, and optionally publishes a signed project receipt. It can also create, inspect, and import offline project bundles. Source audio and model package payloads are excluded by default and require explicit positive MiB limits; project operations are serialized and existing destinations are never replaced.
File and batch jobs keep owner-private recovery records while their exact denoize staging files are live. After a crash, the desktop can retry the saved request or discard the record and only those verified private stage files; existing outputs and batch restart journals are never deleted by recovery. Startup cleanup preserves previews owned by another running desktop instance. The diagnostics export is bounded, owner-private, and no-clobber. It contains only schema-defined capability, limit, recovery-count, and event-code fields; paths, URLs, credentials, device names, free-form errors, and audio are not recorded.
The desktop interface can switch between Japanese and English without a restart and stores only that locale preference. Static interface text and application-owned status messages are checked against the translation catalog during the frontend build. Rust command failures cross the IPC boundary as a stable code, bounded parameters, and a preserved technical detail; the WebView localizes the code instead of treating backend prose as UI copy. Navigation, preview candidates, seek controls, progress, and level meters expose keyboard and ARIA semantics; visible focus, reduced-motion, and forced-colors modes are also supported. CI starts the application in a real Linux WebKit WebView and exercises control names, tab/panel state, keyboard navigation, the skip link, locale changes, live-region semantics, and structured-error IPC.
# Static UI contracts plus the real-WebKit accessibility run
# Build a platform-native installer/package
# Optional FDK-AAC selector
Linux development requires the WebKitGTK 4.1 and GTK 3 development packages. For Ubuntu 24.04 or later:
Prebuilt binaries
Each GitHub Release contains
prebuilt full-feature binaries for:
- Linux x86-64
- macOS Intel and Apple Silicon
- Windows x86-64
The same release also includes CLAP and VST3 plug-in archives for those four
target architectures. Extract the matching denoize-plugin-<tag>-<target>
archive and copy its denoize.clap file or macOS bundle into a standard CLAP
directory. For VST3, extract denoize-vst3-<tag>-<target> and copy
denoize.vst3 into the platform's standard VST3 directory. macOS releases also
contain denoize-auv3-<tag>-<target> for Intel and Apple Silicon; move its
denoize AUv3.app to /Applications and open it once to register the embedded
extension.
Every archive has a matching .sha256 checksum file. Releases also publish the
exact embedded model catalog, its detached signature, the exact embedded model
trust-root document, and denoize-models-<tag>.dmb plus its checksum. The signed
bundle contains the catalog models, upstream licenses, and provenance for
closed-network installation; the standalone catalog/root assets remain
available for independent audit and recovery tooling. Releases also publish
every versioned JSON Schema used by automation, monitoring, DAW state, and
verification clients.
Desktop releases also publish a signed recoverable-update manifest, exact
per-platform SBOMs, two authenticated .dub migration bundles per platform,
and a separate update-asset attestation. A bundle contains both the candidate
and its verified last-known-good installation, so startup-health or explicit
recovery does not require a network. See recoverable application updates.
Every installable CLI archive, CLAP/VST3/AUv3 archive, desktop package, crates.io archive, and offline model bundle also has a per-artifact CycloneDX SBOM. The release evidence archive binds those 24 artifacts and SBOMs to their sizes and SHA-256 digests, while companion GitHub Sigstore/SLSA bundles prove the exact tag commit and release workflow. See release evidence and offline verification for the trust model and an air-gapped verification procedure.
Install with Cargo
The crates.io package provides the CLI and library with every crates.io-compatible backend:
For the embedded DeepFilterNet backend, use a prebuilt GitHub binary or build
this repository with its primary Cargo.toml.
Publishing a release
- Synchronize the root/crates.io manifests and lockfile, the desktop npm and Tauri manifests/configuration and lockfiles, and the generated CLI banner.
- Run
bash scripts/verify-release-version.shto check all 13 version fields. - Commit and push the version change.
- Create the tag from a commit on the default branch and push it:
The GitHub Release workflow verifies that the tag is on the default branch and
matches every release version field, runs the full test suite, and builds all
CLI, CLAP, VST3, and desktop targets before publishing the crates.io package. It then checks
all archives, checksums, signatures, per-artifact SBOMs, build provenance,
recoverable update bundles, and updater metadata before publishing the draft release and generated notes. The
exact .crate archive is attested before publication and its checksum must
match the crates.io API afterward. When docs/releases/vTAG.md exists, its
curated notes are prepended to the generated notes. Desktop startup performs
only local update-health and managed-version repair; an online check, bounded
bundle download, and candidate activation each require an explicit user action.
The legacy latest.json
feed and the recoverable manifest use the same repository signing key. The updater private key is kept in the
TAURI_SIGNING_PRIVATE_KEY repository secret. A failed build leaves the release
as a draft and cannot publish the crate before every target and its evidence
have been verified.
CLI highlights
Portable projects and sample-accurate timelines
Create a source-bound project, review its exact presentation-timeline plan, and assemble it without retaining whole-file PCM:
Selections use exact decoded-presentation ticks, explicit channel maps, silence padding, and only adjacent unpadded crossfades. Manifests bind every source, setting, preset, signed model package, plan, receipt, and license reference by portable locator, byte length, and SHA-256. Changed sources, unsupported graph shapes, unknown fields, path escapes, and output collisions fail before publication. Missing sources can be relocated only to a complete fingerprint and presentation-geometry match.
Offline .dpb bundles carry the manifest, settings, presets, verification
evidence, source licenses, model public keys, plans, and receipts. Source audio
and model package payloads remain references unless their include flag and a
positive aggregate byte ceiling are both supplied. Import authenticates the
complete bundle and publishes only to a new directory. Project batch and watch
automation invoke the same deterministic assembler; watch additionally signs
each successful output.
See portable projects for the CLI, Desktop, bundle, plan, receipt, relocation, and safety contracts.
Read-only plans and signed receipts
Preview exact file, batch, or bounded-stream work without creating output, state, locks, or model-cache updates:
Plans use portable relative locators rather than absolute paths and bind input and model SHA-256 fingerprints, the resolved recipe/backend/accelerator, audio geometry, publication decision, and conservative admitted resources. Planning performs bounded decode, backend/model preparation, and encoder validation, so it can fail before any execution side effect. A skipped batch item also binds the exact fingerprint of the existing output whose journal evidence justified that decision. Stream plans use additive v2 and inspect resumable checkpoint sidecars without locking, repairing, truncating, or deleting them.
Generate an Ed25519 key and publish a receipt only after successful output:
The secret JSON is unencrypted and must remain private. denoize creates it
without clobbering, with Unix owner-only permissions or a protected Windows
DACL, and rejects keys with broader access. A separately supplied public key
or rotation/revocation policy is always required: a receipt never trusts an
embedded signer. Verification authenticates first and then independently
rehashes every rooted output. Batch receipts require the whole batch to finish
without failure or cancellation. Output and receipt are distinct atomic files,
so a final receipt-path race is reported after preserving already committed
audio rather than silently replacing either file. Streaming and stdin receipts
use additive v2 schemas. A captured stdout stream is verified with
receipts verify --output CAPTURED_AUDIO; its receipt is emitted only after
stdout accepts and flushes the complete verified bytes.
The desktop app exposes the same file, batch, and bounded-stream plan preview, plan JSON export, optional receipt publication, owner-private key generation, public-key export, trust-policy creation, and offline receipt/output verification. Secret key paths are session-only UI values and are never stored in desktop settings.
See the stable JSON contracts for schema, privacy, verification, and key-rotation details.
Local authenticated IPC and durable jobs
denoize ipc exposes the same planner, resource admission, atomic publication,
checkpoint, and signed-receipt engine to trusted local automation. It is not a
network service: v1 binds only an ephemeral 127.0.0.1 TCP port, publishes that
endpoint in an owner-private discovery file, and requires an explicit bearer
capability on every framed JSON request. The transport is not encrypted; its
security boundary is the local OS account and the private state/grant files.
Initialize one state directory, then run the foreground server:
The initial administrator capability can manage grants and the server but cannot submit audio. Create a least-privilege worker policy whose canonical input/output roots are disjoint from secrets and state:
Arguments after -- are ordinary processing options. The server rejects flags
that could redirect plans, receipts, resource governors, isolation, model
files, configuration, or publication outside its policy. Dry-run is mandatory
before admission and reports conservative RAM, temporary storage, CPU/GPU work,
destination create/replace/skip counts, overwrite policy, pause support, and an
exact execution-plan digest. V1 executes one job at a time; priority orders the
durable queue and is capped by the submitting capability.
Batch and durable stream jobs pause only after a verified checkpoint or atomic publication boundary and resume by replanning the same request. A daemon crash reclaims them through their lease and checkpoint. A file job has no safe mid-file checkpoint, so an uncertain publication is reported and never retried automatically. Cancellation preserves already published atomic outputs and never emits a false success receipt. A valid signed receipt discovered during recovery wins over an ambiguous process exit. Revocation blocks future requests but does not silently delete already admitted work; use an explicit cancel before revoking when queued/running jobs must stop.
Request/response sizes, request/planning/job timeouts, connections, queue,
history, concurrency, and optional memory/temp/GPU ceilings are finite and
published in denoize-ipc-discovery-v1. Terminal history is bounded and keeps
resource/destination summaries plus plan and receipt fingerprints, not input or
output paths; receipt artifacts are pruned when their history entries age out.
The desktop IPC automation page uses the same Rust client and keeps bearer
tokens outside the WebView. All eight IPC/job schemas ship with releases and
the crate; see the stable JSON contracts.
Realtime audio
Build with the optional system-audio integration, list devices, then route a microphone through a denoising backend to an output or virtual-audio device:
Realtime processing runs outside the device callbacks and uses bounded queues.
The capture callback uses a non-waiting handoff, so an overloaded backend drops
stale chunks; the playback callback emits bounded silence instead of waiting
while the worker publishes a block. --chunk-ms controls the latency/throughput
trade-off and defaults to 100 ms. The low-latency Classical, RNNoise, and causal GTCRN backends are
live-capable; other backend selections are rejected before capture or playback
starts. Input and output devices may use different default sample rates: a
bounded asynchronous sinc converter maps capture frames to the playback clock,
while a PI controller keeps the playback queue near its target without an
abrupt timebase reset. --live-latency 0 selects an automatic target of two
capture chunks with a 40 ms minimum; explicit targets are 20–5,000 ms.
--max-drift-ppm bounds clock correction (2,500 ppm by default, or zero to
disable drift correction while retaining nominal-rate conversion).
Device stream failures enter a finite exponential-backoff recovery loop.
--reconnect-timeout defaults to 30 seconds and zero disables recovery. Named
devices are reselected by an unambiguous exact name; duplicate exact names are
rejected rather than silently routing to a different device. An unspecified
device follows the current system default. Each recovered generation starts
from a cold backend/resampler
state and primes playback before sound resumes. Human-readable status is
written to stderr about once per second. --json instead emits
denoize-cli-output-v1 NDJSON status records containing connection state,
sample rates, queue/latency measurements, drift correction, underruns,
overflows, dropped chunks, reconnects, generation, and accelerator selection.
Latency is an engineering estimate assembled from device callback timing,
capture chunking, algorithmic delay, processing time, and queued playback—not
a loopback measurement or an exact hardware guarantee.
Classical, RNNoise, and GTCRN sessions preserve denoiser, overlap, recurrent, partial-frame, and sample-rate-converter state across consecutive capture chunks. If an overloaded capture queue drops a chunk, the sequence gap clears queued playback and cold resets processing state without reparsing a loaded model before the next retained chunk; state is never shared between separate live sessions. VAD-enabled Classical/RNNoise live processing keeps the legacy chunk-compatible path and prints a one-time warning because causal VAD state and delay alignment are not yet available. GTCRN rejects live VAD rather than silently discarding its recurrent continuity.
Batch processing
Process a directory tree concurrently while preserving its relative layout:
Batch mode validates the complete input/output plan and each input's decoded
audio properties before creating the output directory or starting workers,
then continues after later per-file processing failures and reports a final
summary. Existing outputs remain protected unless --force is supplied, and
input/output directories must not overlap. Recursive discovery does not follow
directory symlinks, and planned destinations that resolve back into the input
tree are rejected.
Omit --output-format only when denoize can re-encode the same container and
codec (WAV, FLAC, Ogg Opus, MP3, AAC-in-MP4, or ADTS AAC). Decode-only
containers such as AIFF/AIFC, CAF, RF64/BWF, plus Ogg Vorbis and ALAC-in-MP4,
require an explicit output format. This prevents implicit Vorbis-to-Opus or
ALAC-to-AAC conversion; for example, use --output-format flac when that
conversion is intentional. AAC-in-MP4 and ADTS AAC also require a build with
the corresponding AAC encoder; unavailable outputs are rejected during
preflight.
With --resume, the v3 state journal binds each completion to the input bytes,
the actual selected backend and effective processing/codec/metadata settings,
any consumed model bytes, the destination identity, and the published output
bytes. A file is skipped only when all of those still match and the output is a
safe regular file with a single link. This also means --backend auto records
the backend it actually selected, not merely the word auto.
Execution-only controls such as memory/temporary/GPU limits, --isolate,
--jobs, and progress output do not change that audio recipe, although their
validation still applies to each run.
Watch-folder automation
Run a durable local inbox that waits for complete files, processes them one at a time, and signs every successful result:
Watch mode polls portably on Linux, macOS, and Windows. A candidate is opened
only when it is a supported regular audio file and its length, modification
stamp, filesystem identity, and SHA-256 content remain unchanged for the full
--settle-ms interval (2 seconds by default). Directory symlinks, FIFOs,
devices, and the separate output tree are never followed as inputs. Use
--once for a bounded settle-and-scan invocation suitable for schedulers; the
default command continues until Ctrl+C. --poll-ms controls the daemon scan
interval, while --max-watch-files bounds each traversal.
Each transition is atomically recorded in
OUTPUT/.denoize-watch-state.json while a sibling lock enforces one writer.
The state binds an opaque digest of the denoize version, processing template,
output format, receipt public-key identity, and any explicitly selected model
or model-key files. Reopening it with a different template fails without
touching prior outputs; choose a fresh --watch-state path to begin a deliberate
new generation.
An interrupted processing entry becomes a due retry on restart. If output
and receipt were both committed before the interruption, their signature,
settled input fingerprint, locator, and output bytes are verified and the job
is recovered without reprocessing. If both disappeared, the same job may be
recreated; a one-sided output/receipt pair is preserved as an operator-visible
failure rather than silently guessed or overwritten. Stable inputs already
recorded as complete are checked with filesystem metadata rather than hashed
on every poll.
Failures use bounded exponential delay from --retry-initial-ms through
--retry-max-ms. --max-attempts defaults to five. A permanent failure or an
exhausted retry budget is copied without clobbering to
OUTPUT/.denoize-quarantine, verified against the settled SHA-256, accompanied
by a denoize-watch-quarantine-v1 JSON explanation, and only then removed from
the inbox. A failed copy leaves the source and a durable quarantine-pending
entry for the next cycle. Custom --quarantine, --receipt-dir, and
--watch-state paths must remain below the output root, which itself must not
overlap the input tree.
The unencrypted receipt key is mandatory and must remain outside both trees and unchanged for the watcher lifetime. A missing or changed key or explicit model artifact defers due jobs without consuming their attempt budgets or quarantining their inputs; restart the watcher with a fresh state path to adopt a deliberate processing-template change.
Outputs preserve relative layout and default to WAV; --output-format selects
another encoder. Existing unrelated destinations are never replaced. When a
later content generation would collide with a prior name, its full content
digest is inserted before the extension.
Watch mode is intentionally sequential and uses the normal per-input resource
governor; --batch, --stream, --resume, --force, --report, --isolate,
and --jobs are rejected.
The desktop Watch folders page uses the same state engine and isolated per-file worker. Select an inbox, a separate output directory, and a receipt secret key outside both trees; then choose the settle/retry policy and start watching. The page displays observed, pending, successful, retrying, quarantined, and superseded counts. Stop prevents another scan and cancels the currently isolated item at its safe publication boundary. Watch paths and the secret-key selection are session-only and are not stored in desktop settings. Processing and resource settings come from the main denoise page, while overwrite remains disabled.
Automatic backend selection
Use --backend auto when the build contains multiple denoisers. Short and
quality-prioritized files use DeepFilterNet when available; long files use
RNNoise to bound processing cost. Realtime sessions prefer RNNoise. The
classical backend is the dependency-free fallback, and the selected backend is
reported before processing.
Adaptive noise profiling
--adaptive-noise detects spectrally noise-like, low-speech-probability regions
throughout a recording and slowly refreshes the classical estimator's anchored
noise profile. This handles changing fans, air conditioning, and room tone
without assuming that the recording begins with silence. Tonal frames are
rejected to reduce the risk of learning sustained notes as noise.
Voice activity detection
--vad detects speech with 20 ms energy frames, hangover, context padding, and
region merging. Long silent spans bypass expensive backend inference and are
strongly attenuated; enhanced speech retains a small dry-signal blend to protect
consonants and attacks. Output channel count and duration remain unchanged.
Loudness delivery
Normalize denoised output to an EBU R128 integrated-loudness target while respecting an oversampled true-peak ceiling:
The applied gain is reduced when necessary to satisfy the peak ceiling, so peak safety takes precedence over reaching the requested LUFS exactly.
Content modes
--mode speech, --mode music, and --mode ambient coordinate related DSP
controls instead of changing only one strength value. Speech mode enables VAD
and adaptive profiling; music mode prioritizes transients, stereo content, and
low suppression; ambient mode preserves environmental texture while tracking
slowly changing noise. Explicit options such as --strength still override the
mode defaults.
Optional FDK-AAC encoder
Pure-Rust oxideav-aac remains the default. Source builders can opt into the
Fraunhofer encoder and select it per invocation:
The FDK feature uses the third-party Rust port and is intentionally excluded
from full and official release binaries. Fraunhofer's codec source has its own
license and MPEG-AAC patent language; downstream distributors are responsible
for reviewing both. The project requires Rust 1.96 or newer.
Raw ADTS AAC
.aac files are decoded and encoded directly as ADTS streams without an MP4
container or an ffmpeg conversion step. M4A and raw AAC share
--m4a-bitrate, which must be a positive kbps value that fits the encoder's
32-bit bps field; raw ADTS output currently uses the default oxideav encoder.
Metadata preservation
File processing merges all readable input tags (for example ID3v2/ID3v1 and APE tags) and remaps the complete set of recognized fields—title, artist, album, track/disc numbers, dates, ReplayGain, lyrics, comments, and artwork—to the destination container's tag type. Cover art bytes, MIME type, picture type, and description are retained by formats that support embedded pictures.
For FLAC and Ogg outputs, arbitrary Vorbis Comment fields are copied verbatim,
including the standard CHAPTER001/CHAPTER001NAME chapter-comment convention.
Their native metadata is scanned and rewritten incrementally instead of
loading the complete audio file into memory. Oversized or malformed input
metadata fails before an output is staged or published.
ID3v2-prefixed FLAC/Ogg files are rejected by this bounded raw path; place
metadata in the container's native comment blocks instead.
When the source and destination use the same native container, format-specific
ID3v2 frames (including CHAP/CTOC) and MP4 atoms are retained as well. A
conversion to a different tag family keeps fields with a defined destination
mapping; container-specific fields without one cannot be represented there.
Use --no-metadata for a clean output.
Quality comparison
|
Quality metrics require sample-aligned PCM. Every input must have a non-zero matching sample rate, the same channel count and frame count, and equal-length channels within each file. Denoize rejects truncated or ragged inputs instead of silently scoring only their common prefix.
The report shows noisy and enhanced SI-SDR, SI-SNR, SNR, segmental SNR, stereo
side SDR, inter-channel correlation error, STOI, PESQ, ViSQOL, and improvement
deltas. It also screens for musical noise, pumping, transient loss, and stereo
phase distortion. These artifact scores are deterministic
dependency-free indicators in [0, 1] (lower is better), not perceptual
listening-test replacements; phase distortion is reported only for stereo
inputs.
When a dB metric is undefined for a silent or otherwise degenerate reference,
the report uses a finite -120 dB floor so JSON output remains valid and
machine-readable.
STOI is calculated natively for sufficiently long reference/test pairs and is
reported in [-1, 1] (higher is better). ViSQOL MOS-LQO is available in the
pure-Rust build when the optional feature is enabled:
|
ViSQOL is a full-reference MOS estimate in [1, 5]. PESQ is intentionally
left as null: the ITU-T P.862 reference implementation and conformance
material require a separately licensed external adapter and are not bundled
with denoize. Inputs that are too short or a disabled optional implementation
are represented as null rather than preventing the rest of the report.
Licensed-corpus release evaluation
denoize evaluate turns quality, output-integrity, and speed claims into one
reproducible, signed release gate. The same strict manifest and runner are
available in the CLI, library, and Desktop app:
Each clean/noisy/model artifact pins a portable path, byte length, SHA-256, SPDX license, source URI and immutable revision, and signal-preparation digest. Audio stays below the caller-selected corpus root and is never embedded in the result or release artifacts. Symlinks, escaping paths, changed files, missing provenance, incomparable PCM geometry, unavailable metrics, and mismatched hardware/runtime/model contexts fail closed.
Signed results include objective and perceptual metrics, duration/rate/channel
agreement, clipping, sample and true peak, DC offset, silence/dropout ratios,
integrated loudness, decode integrity, the canonical output fingerprint,
sorted performance samples, real-time factor, throughput, peak RSS when the OS
exposes it, and every threshold outcome. A required listening protocol cannot
be replaced by automation: run requires a matching, manifest-bound human
result. See reproducible evaluation evidence and the
stable JSON contracts.
Configuration file
Reusable defaults can be stored in TOML and loaded with --config. TOML syntax
and enum names are checked while loading; explicit command-line numeric values
then override file defaults, and the final effective configuration is validated
before input decoding, output staging, or batch worker creation. For example,
FFT frames must be powers of two from 256 through 65,536, streaming blocks must
be from 1 through 1,048,576 frames, batch jobs from 1 through 32, and live
chunks from 10 through 2,000 ms. Non-finite effective floating-point settings
are rejected; live latency accepts zero for automatic or 20 through 5,000 ms,
drift correction accepts 0 through 10,000 ppm, and reconnect timeout accepts 0
through 300,000 ms. Loudness targets are limited to -70..0 LUFS and true-peak
ceilings to -20..0 dBTP.
= "auto"
= "cpu" # cpu|auto|gpu|metal|cuda
= "hifi"
= "speech"
= 0.45
= true
= true
= -16.0
= -1.0
# deterministic = true # serialize processing for reproducible output
# seed = 12345 # optional SGMSE sampler seed (implies deterministic)
# stream_frames = 8192
# chunk_ms = 100
# live_latency_ms = 0 # automatic; otherwise 20..5000
# max_drift_ppm = 2500 # 0..10000
# reconnect_timeout_ms = 30000 # 0 disables hotplug recovery
# max_memory_mb = 1024
# max_process_memory_mb = 2048
# max_temporary_mb = 4096
# max_gpu_memory_mb = 4096
# max_gpu_jobs = 1
# isolate = true # CLI only: run processing in a child boundary
Use --deterministic when an audio result must be reproducible across runs.
The mode serializes channel/model and batch scheduling and uses a stable
stochastic-backend seed. --seed N selects an explicit SGMSE+ seed and implies
the mode. Diagnostic elapsed times and progress messages are intentionally not
part of the reproducibility guarantee.
Batch progress and recovery
Batch runs show completed files, elapsed time, and ETA. --resume records v3
completion entries in .denoize-state under the output directory. The CLI and
desktop app use this same canonical state filename. State written by older v1
or v2 releases is read only as legacy evidence: it is never trusted for a skip.
Run once with --force to regenerate a replaceable legacy output and migrate it
to v3; the following identical run can then skip it.
The resume/force decision is deliberately conservative:
| Planned destination | Without --force |
With --force |
|---|---|---|
| Exact safe v3 output | Skip | Skip |
| Output is missing | Process | Process |
| Legacy, untracked, changed, or unsafe existing output | Error and preserve it | Replace it when the path is safely replaceable |
“Changed” includes input content, the effective backend or recipe, model
content, and output content. Symlinks, multiply linked files, directories, and
special files are never accepted as completed outputs; directories and special
files remain non-replaceable even with --force.
The denoize package version participates in the v3 recipe hash. After a package
upgrade, --resume preserves an existing output and reports recipeChanged
unless --force is supplied. Regenerate it once with --force to migrate the
saved recipe; subsequent identical runs skip it normally.
Resumable ONNX-backed batches require a self-contained .onnx model. ONNX
models that declare external tensor sidecars remain usable in ordinary
non-resume batches, but --resume rejects them because a one-file model digest
cannot safely represent all consumed weights.
Every batch run creates the output directory only after complete input, codec,
and configuration preflight, then takes the shared .denoize-batch.lock before
reading resume state or deciding what to do with destinations. Another denoize
batch using that directory fails immediately while the lock is held. The lock,
canonical .denoize-state, and legacy desktop .denoize-gui-state migration
names are reserved from the planned output topology.
Ctrl+C stops work before publication where possible. Each output is encoded to a randomly named private file in the destination directory. Publication then serializes and synchronizes a journal prepare, the atomic output commit, and a completion record, while rechecking the input and model. Cancellation before this gate leaves the destination and state untouched; once an item enters the gate, its publication is finished atomically. If a process exits between prepare and completion, the next locked batch reconciles the prepared record against the published output before making new decisions. A journal failure closes the gate so later workers cannot publish unrecorded outputs.
Without --force, the destination is checked again at commit time to prevent
ordinary concurrent overwrites. On Unix, the batch output root must be owned by
the current user and must not be group/world writable; output paths through
non-sticky shared-writable or untrusted-owner ancestors are also rejected.
Extended ACLs that grant additional access are also rejected, as are network or
userspace filesystems whose ACLs cannot be verified safely. On Windows, atomic
private staging requires an ACL-capable filesystem such as NTFS; FAT and exFAT
output paths are rejected before encoding, newly created batch control files
receive a protected DACL, and the output root must not be writable by untrusted
accounts. Windows interprocess locking assumes that principals with write or
delete access to the output root or any pre-existing control/output entry
cooperate; denoize does not audit those DACLs for hostile-principal access. Use
--no-progress for quiet
operation or --json for NDJSON progress and summary records. On Unix,
--force also refuses to replace an existing file with an extended ACL or a
different owner, avoiding a silent loss or weakening of its access policy.
JSON summaries retain the cancelled boolean and also report
cancelled_count; together with succeeded, skipped, and failed, that count
partitions the batch total.
These checks define a non-adversarial local-filesystem, process-crash recovery contract. The deterministic resilience matrix exercises every acknowledged journal/checkpoint publication prefix with abrupt child process exits and simulates power loss at local synchronization boundaries. It does not claim protection from a hostile process performing precisely timed ABA path swaps, a lying drive cache, faulty hardware, remote-filesystem semantics, or a kernel/filesystem that violates its documented durability behavior. Keep independent backups for those failure classes.
-b, --backend <NAME> classical|rnnoise|deepfilter
-a, --algorithm <NAME> omlsa|logmmse|mmse|wiener|specsub|specsub-nl|specsub-geo
--window <NAME> hann|hamming|sine|blackman|kaiser|flattop|dpss
--kaiser-beta <B> Finite Kaiser β in 0..50 (default: 8.0)
--dpss-nw <NW> Classical DPSS time-bandwidth product in (0, 8] (default: 3.0)
--multiband Multiband spectral subtraction
--perceptual Bark perceptual gain weighting
--postfilter Musical-noise suppression post-filter
-p hifi Flagship preset (Kaiser + perceptual + postfilter)
--quality ultra Maximum fidelity settings
--onnx-model <PATH> Waveform ONNX model used by the onnx backend
--onnx-rate <HZ> Model sample rate in 1..768000 Hz (default: 16000)
Resilience testing
Every pull request runs fixed parser mutations and deterministic I/O-error and crash-recovery matrices. A scheduled AddressSanitizer/libFuzzer workflow covers all supported audio containers, execution documents, signed receipts and keys, trust policies, and offline model bundles with finite input, RSS, and per-case time limits. See resilience testing for the exact commands, resource-accounting scope, corpus-promotion rule, and debug-only fault protocol.
Library API
use ;
let cfg = HiFi.config;
denoise_file_with_backend?;
// With DeepFilterNet (GitHub/source build with --features full)
denoise_file_with_backend?;
For embedders that use denoise_file_with_backend_config, set
BackendOptions { deterministic: true, ..Default::default() } to serialize
model/channel work. Set seed: Some(value) to reproduce SGMSE+ sampling with
an explicit seed.
For reusable finite-file processing, prepare a common backend session once. Batch workers and VAD regions use this same API, so fixed graphs are shared and dynamic-shape adapters retain their most recent optimized graph:
use ;
let session = prepare?;
let channels = vec!;
let enhanced = session.process?;
assert_eq!;
With the onnx feature, embedders can also inspect and retain the generic
waveform model contract explicitly:
use ;
let model = load?;
assert!;
let channels = vec!;
let enhanced = model.process?;
assert_eq!;
Roadmap status
| Priority | Technology | Status |
|---|---|---|
| 1 | DeepFilterNet v3 | ✅ --features deepfilter |
| 2 | RNNoise | ✅ --features rnnoise |
| 3 | Kaiser/Flat-top/DPSS windows | ✅ |
| 4 | Multiband / nonlinear SpecSub | ✅ |
| 5 | Perceptual weighting + musical-noise PF | ✅ |
| 6 | Pure-Rust external ONNX inference foundation | ✅ validated reusable waveform runtime (--features onnx) |
| 7 | BSRNN / MP-SENet / MossFormer2 adapters | ✅ implemented and quality-gated |
| 8 | SGMSE+ | ✅ 30-step PC sampler + score-model adapter |
See ROADMAP.md for the implementation audit and the acceptance criteria and numerical evidence for each named model.
License
denoize-authored Rust code is MIT licensed. Bundled dependencies and reference materials remain under their respective terms; see THIRD_PARTY.md and LICENSES for notices, corresponding-source information, and license texts.