praatfan-rust

Fast Rust implementation of Praat's core acoustic analysis algorithms, with Python bindings. MIT-licensed.

praatfan_rust is the compiled backend for praatfan — a clean-room reimplementation of Praat's acoustic analysis, written from published papers and Praat's public documentation rather than from Praat's GPL source, so it can be used under a permissive licence.

pip install praatfan-rust

It can be used directly, or installed alongside praatfan and selected automatically through that package's unified API.

What it computes

Audio is read via Symphonia (WAV, FLAC, MP3, OGG, AIFF, CAF and more). NIST SPHERE, including shorten-compressed, decodes through a statically linked desphere — no extra dependency and no external binary. Mono only; multi-channel files require an explicit channel selection.

Example

import praatfan_rust

sound = praatfan_rust.Sound("audio.wav")

pitch = sound.to_pitch_ac(time_step=0.01, pitch_floor=75.0, pitch_ceiling=600.0)
formant = sound.to_formant_burg(time_step=0.0, max_number_of_formants=5,
                                maximum_formant=5500.0)
harmonicity = sound.to_harmonicity_ac(0.01, 75.0, 0.1, 4.5)

Threading

Per-frame analysis (pitch, harmonicity, intensity, formants, spectrogram) and the resampler run on all CPU cores, with results bit-identical to a single-threaded run. Since 0.1.14 results are also the same on every machine: maths functions come from the pure-Rust libm crate and FFTs from a vendored FFTPACK port, so nothing depends on the host's maths library or on CPU features detected at run time. Set RAYON_NUM_THREADS before the first analysis to limit the thread count, e.g. 1 when you already parallelise across processes. The GIL is released during analyses, and the library is safe to use in multiprocessing children started with fork. The WebAssembly build is single-threaded.

Reading harmonicity values

Two entries in an HNR contour are markers rather than measurements:

• -200 — no periodicity was found (silence, DC, unvoiced).
• NaN — the correlation is not resolvable from 1, so no HNR can be derived.

Filter both before averaging, e.g. v[np.isfinite(v) & (v > -200)]. The autocorrelation method also requires periods_per_window >= 3.0; below that the estimator is ill-conditioned and emits a FutureWarning (this will become an error in a future release). 4.5 is the standard value for the AC method; 1.0 is the CC default, not an AC one.

Also available

• WebAssembly — the same core runs in the browser via wasm-pack.
• praatfan-open-pipe — a JSON-over-stdin/stdout binary used by the R package, so R and shell callers pay the audio load once per file.

Links

• Source, documentation and issues: https://github.com/ucpresearch/praatfan-core-clean
• Pure-Python sibling: https://pypi.org/project/praatfan/

Licence

MIT. The wheel also contains third-party code (MIT, Apache-2.0, BSD, Zlib, Unicode-3.0, and the MPL-2.0 symphonia decoders, unmodified); their notices are in the wheel's LICENSE-THIRD-PARTY file.