| name | mir-transcription |
| description | Lead-vocal MIR transcription guidance for SunoScribe. Use when designing, reviewing, or modifying media ingest, vocal separation, RMVPE/F0 extraction, voiced/unvoiced handling, note segmentation, beat/downbeat analysis, quantization, ScoreIR transcription, MIDI export, MusicXML export, or any audio/video-to-score pipeline behavior. |
MIR Transcription
Use this skill to keep SunoScribe transcription work aligned with the MIR principles distilled from doc/principles/mirdoc.md.
Core Rule
Preserve the representation chain:
MediaAsset -> CanonicalAudio -> StemSet -> F0Track -> NoteCandidateSet -> RhythmGrid -> ScoreRevision -> Artifacts
Do not collapse the chain into audio -> MIDI or audio -> MusicXML. Each stage must consume typed input and produce typed output.
Required Workflow
-
Identify the representation being changed.
- Audio waveforms preserve sound but not note labels.
- F0 tracks preserve pitch evidence but not score rhythm.
- MIDI preserves events but not score notation semantics.
- MusicXML preserves notation semantics but is not the analysis source of truth.
-
Keep vocal melody transcription separate from harmony analysis.
- Use vocals for lead F0.
- Use accompaniment/mix later for chord, chroma, structure, or rhythm support.
- Do not use chroma as note transcription; chroma folds octaves.
-
Treat RMVPE/F0 output as evidence, not final notes.
- Preserve continuous F0, voiced/unvoiced state, confidence, and timestamps.
- Detect likely octave errors before quantization.
- Keep slides and vibrato visible in debug artifacts.
-
Build rhythm independently.
- Beat/downbeat analysis must produce a
RhythmGrid.
- Quantization must consume
NoteCandidateSet + RhythmGrid.
- Low downbeat confidence should surface as a warning or issue, not hidden correction.
-
Generate score from ScoreIR or ScoreRevision.
- MIDI and MusicXML must derive from the same semantic revision.
- Do not let export code invent musical decisions that are absent from ScoreIR.
Production Constraints
- Required production stages fail explicitly.
- Missing RMVPE runtime/model is not a successful transcription.
- Vocal separation failure is not a successful lead-vocal MVP.
- Debug artifact failures may warn, but required musical artifacts must not be faked.
- Baseline harmony, structure, or bass guesses must not pollute the lead-vocal MVP.
Debug Artifacts To Preserve
For meaningful diagnosis, prefer generating or consuming:
- waveform
- spectrogram
- vocals/accompaniment previews
- F0 trajectory with voiced/unvoiced regions
- note candidate timeline
- beat/downbeat grid
- quantized note overlay
- MIDI/MusicXML export metadata
Review Checklist
Before accepting a transcription change, confirm:
- The input and output representations are explicit.
- F0 evidence remains available after note segmentation.
- Rhythm grid is independent and inspectable.
- Quantization errors can be traced to F0, beat, downbeat, or note segmentation.
- Exports are generated from ScoreIR, not ad hoc from unrelated intermediate data.
