Multilingual Speech to Text: Test Chinese and English in the Same Sentence
For multilingual speech to text, choose a recognizer that lists each spoken language you need. If you switch between Chinese and English in one sentence, test that sentence before deciding on a model. A language list helps you shortlist models. The test shows whether the workflow keeps both languages, names, and numbers the way you need.
Choose for the way you actually switch languages
| Your task | Start with | What to verify |
|---|---|---|
| One language per paragraph | Set the spoken language for each part when your tool offers that control. | The second paragraph uses the language you selected. |
| Chinese and English in the same sentence | Choose a multilingual model and try a mixed sentence without AI rewriting. | The transcript keeps the words from both languages. |
| Speech in one language, text in another | Transcribe first and apply translation afterward. | The translation preserves the original names, numbers, and conditions. |
For example, the Qwen3-ASR model card lists Chinese, English, and Cantonese, and Whisper offers multilingual variants. In a desktop app, check the exact model variant available to you. Qwen3-ASR model card · Whisper model variants
Test the words you want to keep
Try these samples, or substitute terms you say at work:
- “请把 Roadmap 的链接发给 Mia。” Check the English project name and the person.
- “API gateway 的超时是 30 秒。” Check the technical term and the unit.
- “先别部署，run the smoke test first。” Check that the warning survives and both languages appear.
Keep the microphone, room, and phrases the same when changing models. Save the plain transcript and mark lost words, substituted names, and changed numbers. If you want Traditional Chinese output, check script choice separately from whether the model recognized the right words.
Add cleanup after you check recognition
Try this instruction: “Add punctuation. Keep the original languages, names, terms, and numbers.” Compare the result with the plain version. If a term changes only during cleanup, adjust the instruction or keep the plain transcript. If it is wrong in both versions, test your microphone or a different model.
Voconly · Dictate one sentence in both languages
Choose a downloaded model that supports the languages you speak, test your usual project names, and compare the plain transcript before trying optional cleanup.