Crate cathar 

Audio restoration toolbox — denoise, de-hum, de-click, de-clip, normalise.

Default denoising uses spectral subtraction (pure Rust, zero weights). Enable the ml feature for candle-based neural denoising (Demucs, DNS Challenge).

Quick start

use cathar::{Denoiser, SpectralDenoiser, generate_wave};

let audio = generate_wave(44100, 440.0, 1.0, 0.2);
let denoiser = SpectralDenoiser::default();
let clean = denoiser.denoise(&audio)?;
assert_eq!(clean.channels[0].len(), audio.channels[0].len());

Structs

AudioData

Decoded audio: a sample rate plus one f32 PCM buffer per channel (de-interleaved, sample values in [-1.0, 1.0]).

NoisePrint

Pre-computed noise profile from a silence segment. Feed into SpectralDenoiser::with_noise_print instead of auto-detection.

SpectralDenoiser

STFT spectral-subtraction / Wiener denoiser (see Denoiser).

Spectrogram

A magnitude spectrogram: a column per time frame, each holding bins frequency rows (row 0 = DC, row bins-1 = Nyquist), in decibels.

Enums

Error

Errors returned by cathar’s decode, encode, and processing routines.

Traits

Denoiser

A denoising strategy: turn noisy AudioData into a cleaner copy.

Functions

bandwidth_extend

Restore high-frequency content lost to compression or low sample rates.

breath_remove

Detect and attenuate breath sounds between speech segments.

declick

Detect and interpolate impulse clicks.

declip

Reconstruct clipped samples with A-SPADE sparse declipping (Kitić, Bertin & Gribonval 2015) over a Hann-windowed, 4×-overlapping Gabor tight frame.

deess_multiband

Multiband, adaptive de-esser. The region above crossover_freq is split into bands sub-bands; each tracks its own short-term level (an exponential moving average) and is compressed by ratio only when its instantaneous level rises threshold_db above that adaptive average. Adapting per band and over time catches sibilance concentrated in part of the band and follows a speaker’s changing level, where a single fixed-threshold band over- or under-reacts. Falls back to a single band when bands <= 1.

deesser

Reduce sibilance (harsh “s”, “sh”, “ch” sounds) using HF compression.

dehum

Remove mains hum (50/60 Hz + harmonics) using cascaded notch filters.

deplosive

Tame plosive pops — the low-frequency bursts on “p”/“b” sounds — by attenuating transient energy below ~250 Hz. strength 1–10 (higher removes more). Sustained low-frequency content is preserved.

dereverb

Remove room reverb using spectral envelope decay gating.

derustle

Suppress lavalier / clothing rustle — transient bursts in the ~1.5–6 kHz band are scaled back toward the local temporal median. strength 1–10. Sustained speech in that band is left largely intact.

dewind

Remove low-frequency wind rumble with a 4th-order Butterworth high-pass (two cascaded biquads, ~24 dB/octave). cutoff_hz is the corner frequency (≈ 80 Hz suits most handheld/outdoor wind); content above it is untouched.

dither

Apply triangular-probability-density-function (TPDF) dither at bits resolution. Adds two independent rectangular dithers of amplitude ±0.5 LSB to decorrelate quantisation noise from the signal. Intended for use before truncation to a lower bit depth.

fade

Apply a linear fade-in and/or fade-out to a mono signal. in_sec and out_sec specify the fade durations in seconds. Non-overlapping.

gain_db

Multiply every sample by 10^(db/20). Positive dB boosts, negative cuts.

generate_wave

Generate a mono test tone: a frequency-Hz sine at 0.5 amplitude plus uniform white noise scaled by noise_level, duration_secs long. Uses a fixed seed, so the output is deterministic.

integrated_loudness

Measure integrated loudness in LUFS per ITU-R BS.1770-4 / EBU R128: K-weighting, 400 ms blocks at 75 % overlap, the -70 LUFS absolute gate, then the -10 LU relative gate.

learn_noise_print

Learn a noise profile from an audio segment (should be silence/noise-only).

normalize_peak

Scale to target peak level in dBFS (0 dBFS = ±1.0, -3 dBFS = ~±0.707).

pad

Pad a mono signal with silence at the start and/or end. start_sec and end_sec specify how many seconds of silence to prepend/append.

remix

Remix channels: spec is a list of channel indices or mappings. input_channels are the original channels; output channels are built from spec. Each entry in spec is a list of (channel_index, gain) tuples. The output channel is the sum of input_channel[idx] * gain for each tuple.

resample

Resample one channel from from_rate to to_rate with a Kaiser-windowed sinc (arbitrary ratio). The cutoff tracks the lower of the two Nyquist limits, so downsampling is anti-aliased and upsampling adds no imaging; the filter support widens at low cutoffs to keep the stopband sharp. Returns the input unchanged when the rates already match.

reverse

Reverse a mono signal in time.

select_channels

Select a subset of channels by index.

silence_strip

Strip leading and trailing silence. A sample is “silent” if its magnitude is below threshold_amplitude. Runs shorter than min_duration_sec at the boundary are discarded; gaps shorter than min_duration_sec within non-silent audio are kept.

spectral_repair

Paint out isolated transient spectral artifacts — brief whistles, bursts, and glitches that appear in only a few STFT frames.

spectrogram

Compute the magnitude spectrogram of signal via a Hann-windowed STFT.

trim

Extract a time slice from a mono signal. start and duration are in seconds; samples outside the original range are clamped to the boundary. Returns None if the requested slice is empty.

true_peak_dbtp

Estimate true-peak level in dBTP via 4× polyphase oversampling (the inter-sample-peak method of ITU-R BS.1770-4). Returns f32::NEG_INFINITY for digital silence. Oversampling is fixed at 4×, independent of sample rate.

vad

Simple energy-based voice activity detection: returns (start_sec, end_sec), where the signal rises above threshold_amplitude and stays for at least min_duration_sec. Returns None if no voice segment is found.

variance

Population variance of a sample buffer (mean of squared deviations).

voice_isolate

Isolate speech from background using energy-based VAD + spectral gating.

wiener_denoise

Wiener-filter denoiser — statistically optimal, better transients.