OpenQuack 2.0.0-alpha.18
Pre-release
Highlights
Mixed Chinese/English now holds up across a whole recording. 🌏
Start a thought in English and finish it in 中文 — your Chinese comes back as Chinese, not a botched English "translation," and it's normalised to your system's script (简体 or 繁體).
Before, the streaming engine locked onto the language of your first sentence. Switch to Chinese partway through a long recording and the Chinese tail was silently translated to English:
…continue speaking the next portion entirely in Chinese. I am still trying to read and organize my books with it. The effect is very satisfying…
Now each full chunk re-detects, so the switch is transcribed as Chinese:
…所以,在此刻,让我转过来,并继续在中文中的下一部分说。 最近我还在尝试用它来朗读和整理我的读书笔记…
The script normaliser then puts those characters in your system's script, so you stop getting a random mix of Traditional and Simplified.
Measured on medium (streaming auto, M4 / 16 GB), A/B against the old lock:
| Case | Old (lock) | New (per-chunk) |
|---|---|---|
| Monolingual English (49 s) | WER 3.7% | 3.7% — byte-identical, no regression |
| Monolingual Mandarin (37 s) | hanzi | byte-identical, no regression |
| Code-switch EN→ZH (53 s) | Chinese tail → English | Chinese tail → correct hanzi |
Known limitation: a switch crammed into the final <10 s of a recording lands in the short trailing chunk and is missed (it inherits the previous language). A switch you keep talking in is caught.