@audio/pitch-pyin

import pyin, { track, notes } from '@audio/pitch-pyin'

Mauch & Dixon, 2014. Probabilistic YIN — runs YIN at multiple thresholds weighted by a Beta(2, 18) prior, producing a distribution over candidate pitches instead of a single hard pick. More robust than YIN on ambiguous frames.

let result = pyin(samples, { fs: 44100 })
// → { freq: 440.1, clarity: 0.92, candidates: [{ freq: 440.1, prob: 0.85 }, ...] }

Unlike the other atoms, the single-frame result also carries candidates — the full posterior over detected periods, sorted by probability, each { freq, prob } with prob normalized to sum to 1 across candidates. clarity is the (clamped) total probability mass captured by the candidate set, not a single-peak confidence.

Over a sequence of frames, track runs pYIN's second stage and notes segments its result into notes:

let { times, f0, voiced, prob } = track(samples, { fs: 44100 })   // frame i at i·hopSize/fs, f0 0 where unvoiced
let events = notes(samples, { fs: 44100 })                        // → [{ time, duration, freq, midi, clarity }, ...]

let write = notes({ fs: 44100 })   // streaming, track too
write(block)                       // → the notes finished so far
write()                            // → the rest

track Viterbi-decodes pYIN's pitch HMM. States are 0.1-semitone bins from minFreq to maxFreq, voiced or unvoiced; pitch moves at most 35.92 octaves per second and voicing switches with probability 0.01 (librosa.pyin's defaults). Voiced states take half their candidates' probability (the Vamp plugin's yinTrust), which keeps an unvoiced path open, so a leap between notes is not followed an octave off. A frame is decided once every Viterbi path through it agrees, or, if they have not met within lag frames, along the best path then (fixed-lag decoding, as the pYIN Vamp plugin): memory stays bounded at any length, and a stream with a short lag decides each frame a fixed latency behind its input. Against librosa.pyin 1.0: the difference function is YIN's own (librosa zero-extends the frame, which pulls low pitches sharp), candidates in one bin add up, and f0 is the candidate inside the decoded bin, not the bin centre.

notes is Tony's note model, with the pYIN Vamp plugin's parameters and output: attack, stable and silent states for 3 pitches per semitone. Vibrato and a scoop into a note stay inside it; the same pitch played again starts a new note where the level rises 1/0.7-fold within two frames; a fast glide joins the notes at its ends, a slow one steps through the notes it passes. freq is the median pitch of the note's frames, clarity their mean voiced probability.

candidates replaces stage 1 and leaves the HMM and the note model as they are. It has candidates.js's signature: (frameSize, fs, minFreq, maxFreq) → run, where run(frame) returns the candidate count after filling run.freq, run.prob (each candidate's pitch probability, at most 1 in all) and run.rms (the level the note onsets read); run.frameSize and run.lead set how many samples a frame holds and how many of them precede its time. Without it the output is the same, bit for bit.

Use when: Ambiguous pitched content — breathy vocals, noisy recordings, or when you need a pitch posterior for downstream HMM tracking. Melody of a voice or solo instrument: smoothed f0 with voicing (track), note events (notes).
Not for: Polyphonic audio: one pitch per frame.
Ref: Mauch & Dixon, "pYIN: A Fundamental Frequency Estimator Using Probabilistic Threshold Distributions", ICASSP 2014. Mauch et al., "Computer-aided Melody Note Transcription Using the Tony Software: Accuracy and Efficiency", TENOR 2015.

Part of @audio/pitch — the pitch family umbrella. This README is generated from the umbrella docs.

MIT © audiojs