Skip to content

v0.7.2 — Qwen LiveTranslate Flash + Settings reorg

Latest
Latest

Choose a tag to compare

View all tags
@phuc-nt phuc-nt released this 26 May 13:54
· 3 commits to main since this release
v0.7.2
5efc670

Changes

  • Qwen engine migrated from Omni Plus to LiveTranslate Flash (qwen3-livetranslate-flash-realtime). Server-side VAD replaces the prior RMS-based client VAD; text-only modality; 60+ supported languages via an explicit picker (mirrors mobile v0.4.3). Free preview tier on DashScope international (Singapore region only).
  • Settings reorganized per engine (mobile parity): only the active engine's API key section is shown. Soniox-only features (Custom Context, Strict language detection, Endpoint delay) hide when engine ≠ Soniox. TTS tab hides when engine is cloud-realtime (OpenAI/Qwen). Engine hints collapsed into a single dynamic line.
  • Qwen single-panel transcript: Live Flash returns translation only (no source ASR channel), so the dual-panel view is force-collapsed to single when Qwen is active — no more empty source pane with grey provisional noise.
  • Source picker hides "Auto-detect" on Qwen: Live Flash stalls when source language is "auto" on real mic input. Source language now defaults to English and snaps to a real entry when switching to Qwen.
  • Two-way mode disabled on Qwen (already disabled on OpenAI). Cloud realtime engines are one-way only.

Documentation

  • Installation guides (macOS + Windows, EN + VI) updated with 3 new engine-specific Settings screenshots (setting_soniox.png, setting_openai.png, setting_qwen.png). EN macOS guide gained the previously VI-only Qwen setup section.

Technical

  • src-tauri/src/commands/qwen_realtime.rs rewritten 486 → 264 LOC. URL switched to qwen3-livetranslate-flash-realtime. RMS-VAD state, commit_turn, rms_int16 all removed. New session payload requests modalities=["text"], sets input_audio_transcription.language + translation.language, and disables turn-detection (server handles it). response.text.text snapshots emit as provisional; response.text.done emits final with dedupe on response_id.
  • src/js/qwen-realtime-client.js rewritten 148 → 90 LOC. Provisional buffer is full snapshot assignment (Live Flash sends text + stash every tick, not deltas). connect({ apiKey, sourceLanguage, targetLanguage }) — output queue and source provisional callback dropped.
  • src/js/qwen-langs.js (new): 60-entry language list shared shape with mobile's QWEN_LANGS. No "auto" option.
  • _updateModeUI in src/js/app.js: per-engine visibility for key sections, Soniox-only blocks, and the TTS tab (with active-tab fallback to Translation).
  • TranscriptUI.provider converted to getter/setter so changing provider live-strips the .dual-view CSS class. _render() forces _renderSingle when provider='qwen'. _renderSingle treats Qwen provisional as a target stream (seg-translated class) instead of dim italic.
  • Removed qwen_audio_output from settings (struct, defaults, JS DEFAULT_SETTINGS). #[serde(default)] at struct level handles existing settings.json without migration.

Caveats

  • DashScope keys must be created in Singapore region; other regions hit a different endpoint and fail with WebSocket error on Start.
  • Existing Qwen-Omni users will need to switch source-language picker off "Auto" and may need to re-confirm target language (the auto-snap defaults to Vietnamese if their previous code isn't in the 60-language list).

Full Changelog: ...v0.7.2

Assets 13
Loading