[dependencies]
audio = "^0.9.1"

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

$ audio mp3 speech.wav speech.mp3 --bitrate 128
mp3 — 269184 bytes at 128 kbit/s → speech.mp3

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.

For live audio, vad_stream_new gives a VadStream: push samples as they arrive, poll for closed segments, take their audio. It is a handle — every copy is the same stream, and the last owner releases it; Wavs, segment lists and the encoder's output are plain values. Nothing is released by hand (vad_stream_free / wav_free are optional early releases).

Encoding MP3

mp3_encode is MPEG-1, MPEG-2 and MPEG-2.5 Layer III, in NURL: the polyphase analysis filterbank, the 18-point MDCT with its alias-reduction butterflies, the quantiser's step-size search, and Huffman coding over the big-values regions and the count1 quadruples. All nine sample rates the format defines (32/44.1/48 kHz, 16/22.05/24 kHz, 8/11.025/12 kHz), mono or stereo, constant bitrate. No ffmpeg, no subprocess, no codec library.

There is no psychoacoustic model, and saying so is the honest description of the quality. With no masking threshold there are no scalefactors, so every band gets the same step size, chosen only to fill the frame: the bits go where the signal is loud, not where the ear is deaf. Long blocks only, because deciding when to switch windows is itself a psychoacoustic decision. For speech at 128 kbit/s the difference is not audible; for music at 64 it would be. This is the quality shape of the fixed-point encoder shine, and the comparison below is against it.

The arithmetic is f64 but mirrors a fixed-point encoder's exactly — that is what the three factors of 0.5 in the filterbank and the MDCT are for — so a sample of 1.0 is what a decoder calls full scale and the quantiser lands on the same integers.

frames byte-identical to shine, 17 s of speech	454 / 702
round-trip SNR, this encoder vs shine	18.13 dB vs 18.13 dB
encoding 16.8 s of 24 kHz mono	69 ms

The test needs no decoder. Every frame header says how long its frame is, so a correct file is a chain: the next sync word must land exactly where the previous frame's length says it will. tests/mp3_frames.py walks that chain and fails if one bit was written short or long anywhere in the bitstream — at all nine rates, mono and stereo. When ffmpeg is installed the suite also decodes one back and compares waveforms (correlation 1.0000).

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. A short clip that whisper pads to 30 s pays for far less: log_mel_whisper_padded x 480000 … computes only the frames that see a sample of x — every frame of the padding is the same number — and is bit-identical to padding first (a 2 s clip: ~10 ms).

License

MIT OR Apache-2.0