AI Model Recognition
Use an ONNX model for technique labels in PTECH and UTECH events. Recognition depends on matching the model's labels, input descriptors, and training configuration.
Options
| Keyword | Default | Value | Effect |
|---|---|---|---|
ONNXMODEL | No model | Model file path | Selects the ONNX technique-recognition model. Relative paths are resolved from the score file's directory. |
TIMBREMODEL | No model | Model file path | Alias of ONNXMODEL. Prefer ONNXMODEL in new scores. |
ONNXDESCRIPTORS | Empty list | Space-separated descriptor names | Supplies the model input descriptors in order, unless valid model metadata overrides them. |
Quote paths containing spaces. Model selection and descriptor configuration apply globally when the score is loaded.
The current score grammar accepts these descriptor names:
mfcc logmel loudness rms power chroma zcr hfr centroid spread
flatness flux irregularity kurtosis harmonicity yin
Choose only the descriptors the model expects, in the order used for training. Vector descriptors such as MFCC and chroma contribute multiple input values, so their sizes also matter.
Check the Descriptors For Musicians for a complet overview of the IDs.
Model metadata
Valid, nonempty openscofo.descriptors metadata takes precedence over ONNXDESCRIPTORS. When that metadata is absent or cannot be parsed, the loader uses the score/API list. A model needs a valid descriptor list from one of these sources.
The loader also reads openscofo.sample_rate, openscofo.fft_size, and openscofo.hop_size. Mismatches produce warnings; these metadata fields do not automatically reconfigure the analysis. Match SR, FFTSIZE, HOPSIZE, and the descriptor settings to training.
Labels come from openscofo.labels, with a fallback to classifier labels. Technique names in the score must match available model labels; unknown labels cause score loading to fail.
Example
This example assumes the model was trained at 48000 Hz with a 2048-sample window, a 512-sample hop, the listed descriptors, and labels jet_whistle and pizzicato. Replace these values with your model's actual training setup.
BPM 72
SR 48000
FFTSIZE 2048
HOPSIZE 512
ONNXMODEL "flute-techniques.onnx"
ONNXDESCRIPTORS mfcc logmel centroid flatness hfr flux zcr irregularity kurtosis
MFCCMELS 40
MFCCCOUNT 13
UTECH jet_whistle 1
PTECH pizzicato D4 0.5
See AI Models and Training for creating and exporting models, and Events for technique-event syntax.