Skip to content

Parley 0.11.1

Latest

Choose a tag to compare

@fmasi fmasi released this 11 Nov 20:30
· 4 commits to main since this release
2b32312

Parley 0.11.1 fixes optional echo cancellation, which could leave the first part of a call
uncleaned. It also keeps transcription fully offline even when a model fails to load, and makes
Settings' model update check say something true.

Echo cancellation works from the first minute

With "mic_echo_cancel": "on", the first part of a call on speakers could come out with no echo
removed at all, even though the record said cancellation had run. Parley locked onto a wrong timing
for the echo in the first seconds and kept it. It now measures the timing over the whole part. On a
real speaker call it removed about 49 dB of echo, and on headphones, where there is no echo, it left
your voice untouched. A part where cancellation removed nothing is now recorded as such, not as a
success. (#361)

Transcription never downloads anything

A model that failed to load during transcription used to be deleted and downloaded again on the spot.
That broke Parley's promise that models download only during Setup or when you click Save in Settings,
and left you without a model if the download failed. Now the transcription stops with an error and the
model stays where it is. (#351)

The speech engine's own log lines, which can include recognised words, no longer go to the console.
(#343)

Model updates in Settings

  • Check Now reports an update only when a file Parley actually uses has changed, not on every
    change to the model's page. (#349)
  • Re-download Model in Settings → Transcription fetches a fresh copy. The new copy replaces the
    old one only once it has downloaded completely, so a failed download never leaves you without a
    model. (#349)

Also

  • The command-line transcribe no longer deletes the audio files you give it, and never overwrites an
    existing .m4a. (#362)

Version: a patch release: fixes only, no new capability.