Basic Pitch Accuracy: What Actually Changes the Result

There is no honest accuracy percentage that describes every audio-to-MIDI conversion. A clean piano phrase and a mastered pop song are different inference problems, even when both files use the same sample rate and format.

Basic Pitch estimates notes from spectral and onset evidence. The strongest improvements usually come from making that evidence less ambiguous.

Arrangement matters more than the file extension

A monophonic vocal or lead line gives the model one dominant pitch at a time. Solo piano and guitar are polyphonic, but their note attacks are often distinct enough to produce a useful starting point. Dense mixes are harder because fundamentals, harmonics and percussion overlap.

WAV avoids lossy compression, but it does not remove room noise, clipping or other instruments. A clean MP3 of a solo part can outperform a noisy WAV of a full rehearsal.

The controls solve different errors

Onset threshold changes how much evidence is required to start a note. Lower values recover softer attacks but can add false notes.

Frame threshold changes how confidently a pitch must continue. Raising it removes weak resonance, although sustained quiet notes may become shorter.

Minimum note duration removes events too short to be musically useful. It is especially effective on reverberant guitar, piano pedal noise and compressed recordings.

Frequency limits exclude registers outside the target instrument. They are useful when a bass rumble or bright percussion is not part of the melody you want.

Change one control at a time and compare the note map with the same musical passage. Large simultaneous changes make it difficult to know which adjustment helped.

Sample rate is not a shortcut to accuracy

Tonera prepares a 22,050 Hz mono signal for the model. A 96 kHz upload therefore does not make inference four times more precise. Higher-resolution recording can be useful during production, but isolation, signal-to-noise ratio and clear attacks usually matter more during transcription.

Know when cleanup has reached its limit

Post-processing can remove false positives and reshape timing. It cannot reliably decide which instrument owns an overlapping note. If the target part is buried in a full mix, use stem separation or export the isolated track from the original session.

For a practical test, convert a short representative section first. If the melody and rhythm are recognizable after modest cleanup, the rest of the source is likely worth processing. If every threshold trades one large error for another, improve or isolate the source before spending time editing.