faster-whisper-vocal-score-profile

faster-whisper small.en — Vocal Score Profile

Inference profile — not a trained or fine-tuned model. This repository contains no weights, adapters, recordings, or voice data. It documents the optional lyric-draft configuration used by my Vocal Score research tool with Systran/faster-whisper-small.en.

Role in Vocal Score

Vocal Score turns a monophonic sung performance into editable notation. Melody transcription is handled separately by Basic Pitch. This faster-whisper profile supplies only an optional draft of English words and word timestamps.

The downstream pipeline then assigns a word to a note when their time spans overlap. Notes without a matching word remain blank. Whisper output never changes the detected pitch, rhythm, or score structure.

Profile

The machine-readable settings are in transcription-profile.json.

Output boundary

The raw ASR layer returns timed words. Vocal Score deterministically maps those words onto already-transcribed notes using temporal overlap. word-timestamps.schema.json documents the interchange format.

Intended use

This profile is useful as a small reference for integrating general-purpose ASR into a musician-facing transcription pipeline while keeping melody and lyric responsibilities separate.

Limitations

• Singing is outside the most reliable operating range of ordinary speech recognition.
• Draft lyrics can be wrong, especially with sustained vowels, layered vocals, effects, or a loud accompaniment.
• Word timestamps are approximate and the overlap mapper is deliberately simple.
• No accuracy benchmark is claimed here.
• The draft must be reviewed and corrected by a musician before use.
• No voice cloning, speaker identification, or biometric inference is performed.

License

The original configuration and documentation in this repository are MIT licensed. The upstream model remains governed by its own repository and license.