Faster, smoother transcription and translation — and cleaner text at chunk boundaries.
Speed
- Whisper reads audio in memory instead of writing a temp WAV every pass (skips file write + ffmpeg decode + resample on each transcription).
- Shorter decode fallback on the chunked path caps worst-case re-decoding on hard speech.
- Single resample per pass, shared between the Silero VAD gate and Whisper.
- No re-fed overlap on clean cuts: when VAD cuts on silence, the next chunk starts fresh — less audio per chunk, no boundary duplication.
Translation
- Live streaming translation — the translation appears as it's generated, not all at once.
- Cross-block context — recent source→translation pairs keep terminology, names and pronouns consistent across sentences.
- Ellipsis phrases merge forward — a phrase ending in
…is glued to the next sentence before translating.
Transcription quality (all languages)
- Time-based seam dedup using Whisper word timestamps removes overlap duplicates regardless of spelling (inflection changes, split words).
- Smarter text dedup also collapses short and inflected repeats at chunk seams.
Housekeeping
- The app's boot log is trimmed on each launch so it can't grow without bound.
Full changelog: https://github.com/KazKozDev/live-translation/blob/main/CHANGELOG.md