Skip to content

v0.8.0 — VAD-segmented real-time streaming engine

Choose a tag to compare

View all tags
@apiad apiad released this 29 Jun 13:04
v0.8.0
e31a37f

Highlights

Real-time streaming transcription for long-running audio. Harp now finalizes each spoken chunk at silence boundaries (Silero VAD), decoding it once and never re-transcribing finalized speech — so a long recording streams out as you go instead of waiting for the end. Measured RTF ≈ 0.34 on base/CPU (~3–4× faster than real-time).

Added

  • harp transcribe <file> — stream-transcribe any ffmpeg-readable audio (wav/m4a/mp3/…). -o FILE appends finalized text live; --preview enables a live word-by-word preview; --model/--language select model and language.
  • VAD-segmented streaming engine, FileSource, and harp.vad (SileroDetector, NullDetector).
  • Library-first two-tier TranscriptEvent (committed + transient).

Changed (breaking)

  • CommitEvent → TranscriptEvent; HarpSession gained a detector and streaming knobs.
  • Streaming defaults to beam_size=1 + int8 on CPU; transient preview is opt-in (real-time by default).
  • Step cadence is audio-time-driven, so files stream in chunks instead of one batch decode.

See CHANGELOG.md for details.

Assets 2
Loading