denoize 0.38.0

Pure-Rust audio denoiser with classical DSP and optional RNNoise
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102

# denoize CLI reference

```text
denoize 0.38.0 — pure-Rust audio denoiser engineered for the world's highest sound quality

Classical DSP + optional AI backends (RNNoise, DeepFilterNet v3, MP-SENet, BSRNN).
Input: WAV/BWF/RF64, AIFF, CAF, FLAC, Ogg Opus/Vorbis, MP3, M4A/ALAC, AAC (built in; no ffmpeg).
Output: WAV, FLAC, Ogg Opus, MP3, M4A, AAC.

USAGE:
    denoize <INPUT> <OUTPUT.wav|flac|opus|ogg|mp3|m4a|aac> [OPTIONS]
    denoize live [--input-device NAME] [--output-device NAME] [OPTIONS]
    denoize live --list-devices
    denoize models <COMMAND> [MODEL|all] [OPTIONS]  (run `denoize models --help`)
    denoize metrics <REFERENCE> <TEST> [--json|--markdown]
    denoize compare <CLEAN> <NOISY> <ENHANCED> [--json|--html]

LIVE:
    Low-latency live processing supports only classical and rnnoise; other
    backends are rejected before capture or playback starts.

OPTIONS:
        --config <PATH>      load TOML defaults (CLI options take precedence)
    -b, --backend <NAME>     auto|classical  (default: classical)
    -a, --algorithm <NAME>   omlsa|logmmse|mmse|wiener|specsub|specsub-nl|specsub-geo
    -p, --preset <NAME>      speech|music|aggressive|gentle|restore|hifi
        --mode <NAME>        speech|music|ambient processing intent
    -s, --strength <0..1>    denoising strength (default: 0.6)
        --profile <MS>       finite duration: <0 off, 0 auto, >0 up to 60000
        --no-profile         no profiling; rely on blind IMCRA bootstrap
        --no-adapt           freeze the noise estimate
        --adaptive-noise     learn noise from noise-only regions throughout the file
        --vad                speech-aware segmentation and silence suppression
        --frame <N>          FFT size: power of two in 256..65536 (default: 2048)
        --overlap <F>        overlap ratio 0.5..0.95 (default: 0.75)
        --window <NAME>      hann|hamming|sine|blackman|kaiser|flattop|dpss
        --kaiser-beta <B>    finite Kaiser beta in 0..50 (default: 8.0)
        --dpss-nw <NW>       classical DPSS time-bandwidth product in (0, 8] (default: 3.0)
        --multiband          enable multiband spectral subtraction
        --perceptual         enable Bark-scale perceptual gain weighting
        --postfilter         enable musical-noise suppression post-filter
        --smoothing <0..1>   gain release smoothing (default: 0.6)
        --makeup <DB>        makeup gain in -120..120 dB (default: 0.0)
        --no-dc-block        disable DC-blocking pre-filter
        --quality <LEVEL>    high|ultra
        --no-transient       disable transient/onset protection
        --cepstral           enable cepstral gain smoothing
        --no-cepstral        disable cepstral smoothing
        --pre-emphasis       enable pre/de-emphasis
        --no-pre-emphasis    disable pre-emphasis
        --report             print settings report and exit
        --mp3-bitrate <KBPS> MP3 CBR bitrate (default: 192)
        --m4a-bitrate <KBPS> positive M4A/AAC CBR bitrate (default: 192)
        --aac-encoder <NAME> oxide|fdk (default: oxide)
        --downmix <MODE>     preserve|stereo (default: preserve; lossy outputs reject surround unless explicit)
        --loudness <LUFS>     finite normalization target in -70..0 LUFS
        --true-peak <DBTP>    finite ceiling in -20..0 dBTP with --loudness (default: -1)
        --onnx-model <PATH>   waveform ONNX model (required for -b onnx)
        --onnx-rate <HZ>      model sample rate in 1..768000 Hz (default: 16000)
        --channels <MODE>     independent|linked|mid-side (default: independent)
        --sgmse-profile <P>   fast|balanced|quality (default: balanced)
        --deterministic       serialize processing for reproducible audio output
        --seed <N>            SGMSE sampler seed (implies --deterministic)
        --batch               process files in INPUT directory into OUTPUT directory
        --stream              bounded-memory classical WAV-to-WAV processing
        --stream-frames <N>   block size in 1..1048576 frames (default: 8192)
        --max-memory <MB>     checked working-set limit in MiB (minimum: 1)
        --recursive           include subdirectories in batch mode
        --jobs <N>            workers in 1..32 (default: min(CPU count, 32))
        --output-format <EXT> convert all batch outputs (required when source codec cannot be preserved)
        --force               allow replacing existing output files
        --resume              skip completed files recorded by batch state
        --no-progress         suppress batch progress and ETA output
        --json                emit a machine-readable result
        --no-metadata         do not copy input tags/artwork/chapters to the output
        --input-device <NAME> live capture device (default: system default)
        --output-device <NAME> live playback device (default: system default)
        --chunk-ms <MS>       live chunk duration in 10..2000 ms (default: 100)
    -h, --help               show this help
    -V, --version            show version

BACKENDS (build with --features full for all):
    classical   Enhanced STFT/IMCRA/OMLSA pipeline (default)
    rnnoise     RNNoise via nnnoiseless (requires --features rnnoise)
    deepfilter  DeepFilterNet v3 (requires --features deepfilter)
    onnx        External waveform ONNX model (requires --features onnx)
    mpsenet     MP-SENet magnitude/phase model (requires --features mpsenet)
    bsrnn       ESPnet BSRNN spectral model (requires --features bsrnn)
    mossformer2 ClearerVoice MossFormer2 model (requires --features mossformer2)
    sgmse       SGMSE+ diffusion model (requires --features sgmse)
    gtcrn       Official causal GTCRN for files/library streams (not the live command)

PRESETS:
    hifi        Flagship transparency: OMLSA + protections + advanced DSP
    speech      Voice-optimised balance
    music       Instruments; enables perceptual + postfilter

CONFIGURATION:
    TOML syntax and enum names are checked when loaded. CLI values then override
    TOML numeric defaults, and the final effective configuration is validated
    before audio decoding, output staging, or batch worker creation.
```