PyIN

Pyin, often written pYIN, is a probabilistic pitch-tracking algorithm for monophonic audio. It extends the classic YIN method by combining a frame-level pitch likelihood with a voicing probability, producing an estimate of the fundamental frequency over time along with a confidence or likelihood of voicing for each frame. This combination typically yields more robust pitch contours in modestly noisy signals or passages with vibrato and expressive intonation.

In practice, pyin is implemented in several audio analysis toolkits, with the Python library librosa providing a widely used pyin function. Users feed a mono audio waveform, specify a plausible fundamental frequency range (minimum and maximum F0) and a sampling rate, and receive a sequence of F0 values for voiced frames plus a voicing mask or confidence measure. Frame length and hop size are configurable, affecting resolution and latency.

Applications of pyin include melody extraction, vocal pitch analysis, music transcription, and alignment tasks in musicology and music information retrieval. Its probabilistic framework helps separate voiced from unvoiced segments and can produce smoother, more reliable pitch contours in the presence of noise or expressive singing.

Limitations include a reliance on monophony; polyphonic materials can confuse the estimator. Performance and accuracy depend on parameter choices such as fmin, fmax, frame length, and hop size. Like other pitch detectors, pyin may struggle with abrupt pitch changes, severe noise, or atypical timbres.

Related topics include YIN, probabilistic pitch tracking, and pitch estimation in audio processing.