Skip to content

Parley 0.11.0

Choose a tag to compare

@fmasi fmasi released this 31 Oct 14:46
· 7 commits to main since this release
671fb10

Parley 0.11.0 processes each part of a recording with about a quarter less memory and a quarter less
time. It can now also remove the other side's voice that your microphone picks up from your speakers,
before your words are transcribed.

Faster, lighter processing

For every 30 minutes of recording, Parley now needs about 1.4 GB at its peak instead of 1.9 GB, and
finishes in about 35 seconds instead of 47 (measured on a real call on an M5 Pro). One step read the
whole audio file into memory at full quality before working on it; it now reads it in small pieces.
Your transcript's words and timings are unchanged. (#348)

Optional: speaker detection on the Neural Engine. Adding "diarization_compute_units": "neural_engine"
to ~/Library/Application Support/Parley/config.json brings the peak down to about 0.9 GB, for about
3 seconds more per 30 minutes. It found the same speakers as the default in every test. It is off by
default while it is measured on more Macs.

Optional: remove speaker echo from your microphone

On a call on speakers rather than headphones, your microphone also hears the other side, and that echo
could show up in the transcript as words you never said. With "mic_echo_cancel": "on" in config.json,
Parley removes that echo from a copy of your microphone before transcribing it. (#342, #347)

  • Your recording is untouched. Only the transcription reads the cleaned copy; the audio file and
    the speaker samples are your microphone as it was.
  • It never removes your own words on purpose. Wherever the other side was silent, any of your
    speech that cleaning turned down is put back from the original.
  • It runs on this Mac only, with a small model that ships inside Parley. Nothing is downloaded.
  • Off by default while it is tested on more calls. The echo marking from earlier versions still runs
    on whatever is left.

Also

  • Re-detecting speakers on a transcript whose microphone was cleaned cleans it again, so the speakers
    it finds match the first pass.
  • The command-line transcribe --split --echo-cancel does the same.

Version: a minor release (0.11), not a patch: optional echo cancellation is a new capability.