[dependencies]
audio = "^0.6.0"

audio

Read WAV, resample it honestly, and compute the log-mel spectrogram a speech model actually eats — in pure NURL.

nurlpkg install audio

$ audio info speech.wav
wav — 44100 Hz, 2 ch, 16 bit, 132300 frames (3 s)

$ audio resample speech.wav speech16k.wav --rate 16000

$ audio mel speech16k.wav -o mel.f32
mel — 3000 frames x 80 mels → mel.f32

$ audio vad meeting.wav
59.85 - 71.2 s
190.85 - 202.2 s
vad — 2 segment(s), 7.77 % of the audio is speech

Three places a mistake hides

The WAV reader treats the file as untrusted. Every chunk announces its own size, and a file is free to lie: a data chunk claiming 2 GB behind two bytes, a 0 Hz sample rate, 0 channels, a 12-bit depth nobody has ever used, a header that stops mid-chunk. All of them are a clean error — none of them is an allocation sized by a number a stranger chose. 8-bit unsigned, 16/24/32-bit PCM and 32/64-bit IEEE float all come back as f32 in [-1, 1].

Resampling is windowed-sinc, not linear interpolation. Linear is the tempting one-liner and it is a lousy low-pass filter: downsampling with it folds everything above the new Nyquist frequency back into the speech band — precisely where the model is listening. A 12 kHz tone resampled 44.1 → 16 kHz should vanish, not reappear at 4 kHz. Here it lands at −82.6 dB, and the test suite fails if it climbs above −60.

The mel spectrogram is whisper's, to the constant. Every one of these was read out of transformers' source, not remembered — and every one of them silently changes what the model hears:

window	periodic Hann, hanning(N+1)[:-1] — the symmetric one is a different window
padding	reflect, by n_fft/2 — so frame k is centred on sample k·hop, not started there
mel scale	Slaney (linear below 1 kHz, log above) — not HTK's 2595·log10(1+f/700)
filters	Slaney-normalised, 2/(f[k+2] − f[k])
after	drop the last frame, log10, floor at max − 8, then (x + 4)/4

Verified against the thing itself

Not against my own understanding of it:

• log-mel == Hugging Face's WhisperFeatureExtractor, on real audio:

max |Δ| = 1.8e-5, mean |Δ| = 8.0e-9, correlation 1.00000000 over all 80 × 3000 values. (The residual is f32 storage rounding.)

• the resampler tracks scipy's resample_poly at r = 0.99992, and kills

the out-of-band tone by 82 dB

• an independent numpy reference (tests/mel_ref.py) guards against

regressions when transformers is not installed

• every bit depth round-trips; six malformed files are six clean errors;

ASan/LSan clean

Finding the speech

A recording is mostly not speech, and a speech model does not get cheaper by running over silence — it costs exactly the same. vad_segments says where the speech is.

It is an energy detector, and being clear about that matters, because it decides where it fails: it hears "something loud enough, often enough" rather than "a human voice", and it will call a slammed door speech. faster-whisper reaches for Silero, a small neural VAD, for exactly that reason. This is the honest version of what can be done without a second model, and the seam is here for one.

What it is not is a fixed threshold. A fixed dB threshold works on the file it was tuned on and nowhere else. The floor here is the 10th percentile of this recording's own frame energies — whatever the quiet part of this room sounds like — and speech is what stands 6 dB above it for 250 ms. Gaps under 400 ms are bridged (that is a pause between words, not between sentences) and 200 ms is kept either side, because speech starts before it gets loud and a word's tail is quiet.

vad_extract gives back the audio with the silence taken out — keeping at most max_gap samples of the real room between segments. That cap is not a detail: glue two utterances together with nothing between them and a speech model hears one utterance where there were two.

vad_extract_runs additionally hands back the condensation map — one VadRun { cond, orig, len } per surviving stretch — because a model transcribing the condensed audio reports times in the condensed timeline, and a caller who wants to say "this was said at 3:12 of the recording" has to walk the map back. vad_map_sample does.

Built on

stdlib/std/fft.nu. Padding 400 to 512 does not compute a rounder spectrum; it computes a different one — so 400 is transformed exactly, and how it is transformed is picked by what 400 actually is: 2^4·5^2 is perfectly factorable, so it takes the mixed-radix Cooley-Tukey path (Bluestein, the chirp-z fallback, is for lengths with a large prime factor — a 997-point frame would take it; a 400-point one paying for two 1024-point transforms to get one 400-point answer was the wrong deal). A real signal of even length is also folded first: 400 real samples become a 200-point complex transform plus an O(n) untangle, and half the transform is half the work.

The whole 30-second mel spectrogram (3000 frames) takes about 0.14 s, down from 0.52 when every frame went through Bluestein at full length.

License

MIT OR Apache-2.0