v3.1.0 — 5 new Nova-3 languages (af, hy, ka, ne, pa), 54 translation targets + stable v3

First stable release of the v3 line — includes everything from the v3.0.0 changelog entry (previously only available as v3.0.0-beta.1) plus five new languages.

New in 3.1.0

Five new Nova-3 transcription languages — Afrikaans (af), Armenian (hy), Georgian (ka), Nepali (ne), and Punjabi (pa), following Deepgram's August 2026 Nova-3 expansion (batch and streaming). Each gets a language-tagged sidecar (.afr.srt, .arm.srt, .geo.srt, .nep.srt, .pan.srt), audio-stream matching, and a web-UI dropdown entry. All five are also available as LLM translation targets, bringing translation coverage from 49 to 54 languages.

Verified end-to-end against the live Deepgram API with native-speaker recordings: transcription (af, ka), non-Latin keyterm prompting (Georgian script), and translation into the new targets (pa, ne).

Also new since v2.4.1 (from 3.0.0)

Local transcription engine (Whisper, beta) — opt-in -local Docker images bundle faster-whisper (CPU, int8) so the whole pipeline (transcription, keyterms, translation via Ollama) can run offline at $0 with nothing leaving your machine. Deepgram Nova-3 remains the default engine; select per job in the Web UI or via ASR_ENGINE=whisper. Cloud-engine subtitles no longer embed [speaker N] tags by default (re-enable with the "Speaker labels in subtitles" toggle or SPEAKER_LABELS=1).

See CHANGELOG.md for the full 3.0.0 fix list.