MuScriptor

MuScriptor is an audio-to-symbolic tool that converts music audio into note-event JSON or a MIDI file. The default download is the standalone F32 GGUF package; the original safetensors layout is still supported for local development.

python3 tools/model_manager_v2.py install muscriptor

audiocpp_cli --task midi --family muscriptor \
  --model models/MuScriptor-Small-GGUF/muscriptor-small-f32.gguf \
  --backend cuda \
  --audio song.wav \
  --request-option instruments=drums,electric_bass \
  --out result.mid

Streaming mode accepts audio chunks and returns the final MIDI/event result:

audiocpp_cli --task midi --mode streaming --family muscriptor \
  --model models/MuScriptor-Small-GGUF/muscriptor-small-f32.gguf \
  --backend cuda \
  --audio song.wav \
  --request-option instruments=drums,electric_bass \
  --out result.mid

Option	Values	Default	Meaning
--request-option instruments=<csv>	instrument names	empty	Constrain generated MIDI events to instrument groups.
--request-option output_format=midi|json	midi, json	midi	Select the primary output serialization written by --out.
--out <file>	path	not set	Write the selected primary output. Use .mid for MIDI or .json for event JSON.
--text-out <json>	path	not set	Optionally also write generated note-event JSON.
--max-tokens / --request-option max_tokens=<n>	integer	2000	Maximum generated MIDI-event tokens per chunk.
--request-option do_sample=true|false	bool	false	Use sampling instead of greedy token selection.
--temperature	float	1.0	Sampling temperature. 0 selects deterministic argmax.
--guidance-scale / --request-option guidance_scale=<f>	float	1.0	Classifier-free guidance coefficient; 1 disables CFG.
--num-beams / --request-option num_beams=<n>	integer	1	Beam-search width; 1 disables beam search.
--request-option batch_size=<n>	integer	1	Number of chunks decoded per batch when prelude_forcing=false.
--request-option prelude_forcing=true|false	bool	true	Force open-note prelude tokens between sequential chunks.
--seed	integer	0	Sampling seed.
--session-option muscriptor.weight_type=<type>	native, f32, f16, bf16, q8_0	native	Transformer weight storage type.
--session-option muscriptor.perf_mode=<mode>	off, flash_attention	flash_attention	Decoder attention mode.
--session-option muscriptor.weight_context_mb=<mb>	integer MiB	512	Weight context arena size.
--session-option muscriptor.conditioning_graph_arena_mb=<mb>	integer MiB	128	Condition graph arena size.
--session-option muscriptor.decoder_prefill_graph_arena_mb=<mb>	integer MiB	768	Decoder prefill graph arena size.
--session-option muscriptor.decoder_decode_graph_arena_mb=<mb>	integer MiB	512	Decoder cached-step graph arena size.