-
-
Notifications
You must be signed in to change notification settings - Fork 0
Glossary
Justagwas edited this page Jul 15, 2026
·
1 revision
| Term | Meaning in A2M |
|---|---|
| Attribute model | ONNX component that predicts velocity and refined onset/offset attributes for decoded intervals |
| Batch | Number of overlapping audio sections sent through the scorer together |
| CPU provider | ONNX Runtime execution on the processor |
| CUDA | NVIDIA GPU execution provider used by the managed CUDA runtime |
| DirectML | Windows GPU execution provider used by A2M's DirectML runtime |
| F1 | Harmonic mean of precision and recall; balances extra and missed events |
| Frontend | Analysis windows and mel filter data used to turn waveform frames into model input |
| Hop | Sample or time step between adjacent analysis frames or sections |
| Inference | Running model inputs through an ONNX model to obtain outputs |
| Mel feature | Frequency representation grouped on a perceptually motivated scale |
| MIDI | Symbolic event format containing notes, timing, velocity, and controllers rather than sound |
| Note onset | Start time of a note |
| Note offset | End time of a note |
| ONNX Runtime | Runtime used to execute A2M's scorer and attribute ONNX models |
| Precision | Fraction of matched output events among all output events under a stated matching rule |
| Provider | ONNX Runtime backend such as CPU, CUDA, or DirectML |
| Recall | Fraction of reference events matched by output events under a stated matching rule |
| Runtime pack | Provider-specific ONNX Runtime files managed separately from the main app |
| Scorer | ONNX component that scores possible event intervals, skip/noise paths, and context |
| Section | Overlapping portion of the recording processed by the engine |
| Velocity | MIDI value from 1 to 127 representing note attack strength in the output |
| Viterbi decoding | Dynamic-programming search that selects a high-scoring sequence of event intervals |