kizuna-ai-lab/sokuji v0.25.2

on GitHub

What's new in v0.25.2

A focused fix for the karaoke-style highlight on the OpenAI Translate provider.

🐛 Fixes

• Karaoke highlight no longer snaps backward mid-sentence on OpenAI Translate (#216). On longer translated utterances the in-bubble highlight used to repeatedly jump back to the start of the sentence before catching up, especially around natural pauses between words. It now advances smoothly with the played audio and reaches the end of each sentence cleanly, instead of stalling a couple of characters short and then jumping to the period.

🔧 Internal

• OpenAI Translate now anchors per-chunk audioSegments on each session.output_audio.delta so the karaoke renderer steps the highlight in chunk-aligned units instead of the proportional time fallback.
• The translate client reads sample_rate from the event with a fallback to the spec default, in case the API ships a different rate in the future.

Full Changelog: v0.25.1...v0.25.2