Crate denoize 

denoize — pure-Rust audio denoiser built for the world’s highest fidelity.

Goal: transparent, artifact-free restoration that preserves timbre, transients, dynamics, and “air” better than any classical offline tool.

Implemented technologies

Classical DSP (always available)

• STFT/ISTFT + Perfect Reconstruction OLA（高オーバーラップ対応）
• IMCRA/MCRA ノイズ推定 + Spectral Flatness プロファイル + Anchoring
• Ephraim-Malah Decision-Directed SNR
• 8種類のゲイン推定器（OMLSA, LogMMSE, MMSE-STSA, Wiener, SpecSub + 非線形/幾何学的）
• Attack/Release + Cepstral Smoothing + Transient Protection
• 高度窓関数: Kaiser / Flat-top / DPSS
• マルチバンドスペクトルサブトラクション
• 知覚重み付け（Bark帯域）+ 音楽ノイズ抑制ポストフィルタ

Input / output codecs (built-in, no ffmpeg)

• Decode: WAV / MP3 (symphonia, including Xing/LAME gapless timing, with a bounded raw-stream compatibility fallback) / M4A (Pure-Rust AAC-LC and ALAC with v0/v1 unity-rate edit-list presentation timing)
• Encode: WAV / MP3 (shine-rs) / M4A (oxideav-aac Pure-Rust AAC-LC)
• Decoded to f64 PCM at native sample rate (no extra quantisation)

Optional AI backends (feature-gated)

• rnnoise feature: RNNoise via nnnoiseless (pure-Rust)
• deepfilter feature: DeepFilterNet v3 via tract ONNX
• onnx feature: user-supplied waveform ONNX models via tract (including a reusable, contract-checked loaded-model API)
• mpsenet feature: MP-SENet compressed-magnitude/phase ONNX adapter
• bsrnn feature: ESPnet BSRNN spectral ONNX adapter
• mossformer2 feature: ClearerVoice MossFormer2 48 kHz ONNX adapter
• sgmse feature: SGMSE+ iterative diffusion ONNX adapter

Build with all backends: cargo build --release --features full

Re-exports

pub use atomic_output::AtomicOutput;

pub use atomic_output::CommitMode;

pub use audio::ensure_memory_limit;

pub use audio::estimate_audio_memory_bytes;

pub use audio::estimate_audio_working_set_bytes;

pub use audio::estimate_file_memory_bytes;

pub use audio::estimate_session_memory_bytes;

pub use audio::estimate_stream_memory_bytes;

pub use audio::estimate_stream_memory_bytes_checked;

pub use audio::inspect_wav_session;

pub use audio::read_audio;

pub use audio::read_audio_from_session;

pub use audio::read_audio_from_session_with_limits;

pub use audio::read_audio_with_limits;

pub use audio::read_audio_with_metadata_limits;

pub use audio::read_wav;

pub use audio::read_wav_bytes;

pub use audio::read_wav_bytes_with_limits;

pub use audio::read_wav_from_session;

pub use audio::read_wav_from_session_with_limits;

pub use audio::read_wav_with_limits;

pub use audio::sanitize_sample;

pub use audio::write_audio;

pub use audio::write_wav;

pub use audio::write_wav_bytes;

pub use audio::write_wav_channel_mask;

pub use audio::Audio;

pub use audio::WavStreamInfo;

pub use audio::WavStreamReader;

pub use audio::WavStreamWriter;

pub use backend::gtcrn::GtcrnModel;

pub use backend::gtcrn::GtcrnStream;

pub use backend::onnx::OnnxWaveformContract;

pub use backend::onnx::OnnxWaveformLayout;

pub use backend::onnx::OnnxWaveformModel;

pub use backend::decode_mid_side;

pub use backend::encode_mid_side;

pub use backend::Backend;

pub use backend::BackendOptions;

pub use backend::BackendSession;

pub use backend::ChannelMode;

pub use backend::OnnxModelConfig;

pub use backend::SgmseProfile;

pub use backend::StreamingBackendSession;

pub use benchmark::ArtifactReport;

pub use benchmark::BenchmarkReport;

pub use benchmark::ComparisonReport;

pub use channel_layout::ChannelLayout;

pub use channel_layout::ChannelMask;

pub use channel_layout::ChannelPosition;

pub use channel_layout::PanInfo;

pub use config::ConfigError;

pub use config::ResourcePlan;

pub use decode::decode_file;

pub use decode::decode_file_from_session_with_limits;

pub use decode::decode_file_with_limits;

pub use decode::decode_file_with_metadata_limits;

pub use decode::probe_file;

pub use decode::probe_file_from_session_with_limits;

pub use decode::probe_file_with_limits;

pub use decode::probe_file_with_metadata_limits;

pub use decode::AudioCodec;

pub use decode::AudioFormat;

pub use decode::AudioProbe;

pub use decode::DecodeLimits;

pub use decode::DecodedPcm;

pub use denoiser::Denoiser;

pub use denoiser::DenoiserConfig;

pub use denoiser::Preset;

pub use denoiser::ProcessingMode;

pub use denoiser::StreamingDenoiser;

pub use encode::AacEncoder;

pub use encode::DownmixMode;

pub use encode::EncodeOptions;

pub use encode::OutputFormat;

pub use gain::Algorithm;

pub use gain::SpecSubLaw;

pub use input::AudioInputSession;

pub use quality::QualityMetrics;

pub use window::WindowParams;

pub use window::WindowType;

Modules

atomic_output

Atomic creation and replacement of filesystem outputs.

audio

Audio file I/O: WAV read/write (hound) + unified decode for MP3/M4A/WAV.

backend

Optional AI denoising backends (feature-gated).

benchmark

Objective quality and stereo-imaging reports.

bessel

Special functions needed by the Ephraim-Malah family of gain estimators: modified Bessel functions of the first kind, orders 0 and 1, and the exponential integral E1.

channel_layout

Standard channel layouts used by the audio pipeline.

config

Shared validation and resource-planning primitives for external input.

decode

denoize 自作デコード層 — lossless / lossy audio containers を高品質 PCM (f64) へ。

denoiser

The de-noizer: ties together STFT, IMCRA noise estimation, the decision-directed a-priori SNR estimator, the selected spectral gain estimator, attack/release + cepstral gain smoothing, transient protection, and optional pre-emphasis.

encode

Audio encode layer — WAV / MP3 / M4A output.

fft

A self-contained, allocation-light, iterative radix-2 Cooley-Tukey FFT.

gain

Spectral gain estimators.

input

Stable, regular-file-backed audio input sessions.

loudness

EBU R128 / ITU-R BS.1770 loudness measurement and normalization.

metadata

Cross-container audio metadata preservation.

models

Signed external-model catalog, verified local cache, and installation provenance.

noise

Noise-power-spectral-density estimation.

perceptual

Perceptual weighting for spectral gain control.

postfilter

Musical-noise suppression post-filter.

quality

Full-reference perceptual quality metrics.

resample

Band-limited, channel-synchronous sample-rate conversion.

service

Shared application service used by the CLI and graphical frontends.

stft

Short-time Fourier transform engine with perfect-reconstruction overlap-add (OLA) synthesis.

stream

Stateful block-processing API for realtime and pipe integrations.

vad

Lightweight energy VAD and speech-region segmentation.

window

Window functions for short-time Fourier analysis/synthesis.

Functions

denoise_audio_with_backend_config

Process already-decoded audio in place. This is the path used by stdin and embedders that do not have filesystem-backed input.

denoise_audio_with_backend_session

Process decoded audio with a prepared backend graph.

denoise_file

Denoise a WAV file end-to-end, writing the result to output.

denoise_file_with_backend

Denoise with an explicit backend (classical / rnnoise / deepfilter).

denoise_file_with_backend_config

Denoise with explicit backend, encoder, and backend-specific model options.

denoise_file_with_backend_opts

Denoise with explicit backend and output encode options.

write_audio_transactional

Encode audio and optional metadata into a staged file, then publish it in one filesystem commit.

write_audio_transactional_as

Encode audio using a format selected during preflight, then publish it in one filesystem commit without re-inferring the codec from the path.