MuScriptor.Stream

Name	Mandatory	Description	Default	Type
⬅️ Input		The input of the shard, if any		[Float]
Output ➡️		The resulting output of the shard		[{pitch: Int onset: Float offset: Float program: Int drum: Bool instrument: String}]
Model	No	The MuScriptor model to use.	none	Var(MuScriptorModel)
SampleRate	No	Sample rate of the input samples in Hz; each segment is resampled to 16000 if different.	16000	Int
Instruments	No	Instrument names to condition the model on (e.g. "acoustic_piano" or ["acoustic_piano" "drums"]).	none	NoneStringVar(String)[String]Var([String])
MaxTokens	No	Maximum generated tokens per chunk.	2000	Int
Temperature	No	Sampling temperature. 0 = greedy decoding (the reference default).	0	Float
TopP	No	Top-p (nucleus) sampling; takes precedence over TopK. 0 disables.	0	Float
TopK	No	Top-k sampling. 0 disables.	0	Int
Seed	No	Random seed used when Temperature > 0.	299792458	Int
Drain	No	When true, after consuming this activation's input, zero-pad and transcribe the remaining partial segment, close all open notes and reset the stream state.	false	BoolVar(Bool)
Reset	No	When true, discard all buffered samples and decoder state before consuming this activation's input.	false	BoolVar(Bool)

Streaming audio-to-MIDI transcription. Feed mono samples incrementally (e.g. from a live audio wire); the shard buffers them, transcribes each full 5-second segment as it becomes available and keeps the decoder state across activations so notes sustained across segment boundaries survive. Outputs the notes completed during this activation (often empty while buffering). Set Drain: true on the final call to flush the remaining partial segment and reset for a new stream.

---