-
-
Notifications
You must be signed in to change notification settings - Fork 0
FAQ
The current engine is structured around all 88 piano pitches plus sustain and soft-pedal events. Other sounds may produce output, but that is not broad instrument support.
No. Transcription runs locally. A2M can still use the internet for its model, optional GPU components, update checks, or an update download. See Privacy, Networking and Security.
No. CPU is the normal starting point. CUDA and DirectML are optional execution providers for the same transcription engine.
It does not select a different model. Provider-specific floating-point differences are possible, but GPU mode is an execution choice, not a more accurate engine.
Processing time depends on recording length, provider, settings, session startup, CPU, GPU, drivers, and batching. Test representative files on the computer that will run A2M rather than relying on a universal ratio.
There is no honest single percentage for every piano, room, microphone, mix, or musical style. See Worked Example and Validation for one fully inspectable input/output example and exact metric definitions, and Accuracy, Limitations and Best Results for practical interpretation.
A note attack can create a distinct transient. Its release may be obscured by resonance, sustain pedal, reverb, and nearby notes. Onset and offset quality should therefore be inspected separately.
Expressive mode writes predicted note velocities. Uniform mode replaces them with one chosen value. Pedal export can be disabled; when enabled, decoded sustain and soft-pedal symbols are written as CC64 and CC67. The current model does not produce sostenuto CC66.
A2M preserves an existing MIDI file by choosing the next unused suffix.
Yes. Cancellation is cooperative and checked at safe points, so the current operation may take a moment to return.