github kizuna-ai-lab/sokuji v0.35.1

5 hours ago

Highlights

Soniox

  • Audition a cloned voice without starting a session. After cloning, ready voices in "Manage imported voices" get a play button — one click synthesizes a short sample so you can judge the clone, and re-record the reference clip if it came out poorly, without spinning up a full translation session. Switching voices or closing the panel cancels the request in flight. (#379)

OpenAI

  • Two new transcription models. gpt-live-transcribe and gpt-transcribe are now selectable for your own transcript. The default stays on gpt-4o-mini-transcribe, which remains the cheapest option. (#378)
  • Your source language now reaches the transcriber. It was configured in settings but never actually sent, so speech recognition had no language hint to work with. (#378)
  • Transcription keywords. A new glossary field lets you prime names, jargon, and product terms so they come through correctly. Terms are hints — one only appears in the transcript when it is actually spoken. Available with gpt-live-transcribe and gpt-transcribe. (#378)
  • OpenAI Translate moved to gpt-live-transcribe for the source transcript — same cost as the legacy model it replaces, with a lower word error rate. Existing settings migrate automatically. (#378)
  • Fixed: in Participant and Both modes the transcription language hint pointed at the wrong side of the conversation, working against recognition of the other party's speech instead of helping it. (#378)

These transcription changes affect the original text shown beside each translation. The translation itself is produced from your audio directly and is unchanged.

Full changelog: v0.35.0...v0.35.1

Don't miss a new sokuji release

NewReleases is sending notifications on new releases.