PhonePaint

Correct mispronounced phones in a recording, or play around editing speech phone by phone. This tool is especially useful for removing lateralized pronunciation and pharyngeal compensatory effects, replacing each affected phone with itself (e.g. s --> s), see examples on https://storage.googleapis.com/phone-paint-demo/demo.html.

1. Analyze

Aligner

MAPS runs in-process. MFA is usually more accurate but needs a conda Kaldi sidecar, which a Space image cannot install.

Median boundaries across 10 checkpoints: steadier, slower.

Example clip (VCTK p225)

Aligned phones

Forced alignment by MAPS. Audio codec weights are VoiceCraft's 16 kHz EnCodec, licensed for non-commercial use only.