kizuna-ai-lab/sokuji v0.19.0 on GitHub

What's New

Edit the system prompt your local translation model receives. Qwen-family models now get a dedicated Translation Prompt section in settings with two modes:
- Simple — a read-only preview of the improved default prompt.
- Advanced — two freeform textareas (one for you, one for other participants) where you can write your own instructions.
Fixes the "chat instead of translate" problem. Small models like Qwen3 0.6B sometimes misread Chinese input like "我想问一些问题" as a question and respond "Sure, how can I help you?" instead of translating. You can now override the prompt to enforce strict translation behavior.
Works per-direction. Your prompt applies to your speaking direction; the participant prompt applies to the reverse direction (which may use a different Qwen model). Non-Qwen translation models (Opus-MT, TranslateGemma) can't consume freeform prompts and show the section as disabled with an explanation.

Three new library ID fields for the Volcengine self-learning platform, letting you bias recognition and translation toward your own terminology:
- Hot words — improves ASR accuracy for proper nouns and domain-specific terms.
- Replacement — post-processing text replacements on recognized speech.
- Glossary — source→target term pairs that guide translation output.
Paste the library IDs from your Volcengine account into the new Custom Vocabulary section; they're passed on every recognition request.

Slack-style conversation rows — per-speaker avatars and filter buttons make it easier to follow who said what in multi-participant sessions.
Compact conversation mode — a denser layout option for when you want to see more history at once without scrolling.
Role-colored translation badges — translations from participants are highlighted in a distinct color (orange) so you can tell at a glance which side of the conversation each line came from.

Full Changelog: v0.18.1...v0.19.0