How accurate is Chinese speech to text with local Whisper?
In the Whisper paper itself (arXiv:2212.04356), the large-v2 model scores 14.7% character error rate on Chinese FLEURS. That is a character error rate rather than a word error rate: Chinese is written without spaces, so there is no word boundary to score against, and the paper uses character-level scoring for Chinese, Japanese and Korean for that reason. A Chinese figure and a Spanish word error rate are not the same measurement.
As always that is clean read speech and describes a floor rather than an expectation. Real dictation in a normal room runs higher. Accuracy tracks model size far more than it tracks local versus cloud - it is the same open-source Whisper either way - and the simplified-versus-traditional problem below is noticeably worse on the compact models.
The honest headline for Chinese is that most of the public frustration with Whisper is not about how well it hears you. It is about which characters come out, which is the next section.
The simplified versus traditional problem, and why we do not solve it
Whisper emits simplified or traditional characters unpredictably, on the same audio. This is the most frequently reported Chinese issue across Whisper's own discussions and downstream projects - users in mainland China expect simplified output and get traditional, or get a mixture within one transcript. The reported cause is the proportion of traditional-script text in the training data, much of it from subtitles, and the effect is strongest on the smaller models.
We have no fix, and we will not pretend otherwise. The two known workarounds are a Chinese initial prompt, which does not reliably work, and post-processing through a converter such as OpenCC. SnailText ships neither. It is also worth knowing that conversion is not a safe mechanical step: the mapping between the two scripts is not one-to-one, and some cases require resolving homophones from context, so an automatic converter can introduce its own errors.
One language code covers every variety of Chinese. Whisper uses a single code for all varieties, and its maintainers have said so directly. You cannot use the language setting to ask for a particular script, a particular dialect or a particular regional standard. We do not claim support for Cantonese, Hokkien or Shanghainese - the code is the same one, and we have not measured quality on non-Mandarin varieties.
What we actually offer for Chinese, and what we do not
For Spanish, German, French, Portuguese and Dutch, SnailText ships a language-specific prompt. For Chinese we do not. No Chinese prompt, no cleanup pass, no character converter, no script preference, and no claim about tone handling - we found no measurement of how much tone confusion contributes to Whisper's Chinese errors, so we are not going to assert one.
What SnailText does give you is worth being concrete about, because for Chinese the speed argument is stronger than for most languages. Typing Chinese means a two-stage pinyin process: type the sound, then pick the character from a candidate list. Speaking skips both stages. With a global hotkey the text lands at your cursor in any app, and the audio is processed in RAM and never uploaded.
That makes this a reasonable tool if you want fast local Chinese dictation and can live with fixing the character script yourself, and a poor one if guaranteed simplified output is a requirement. We would rather you knew which of those you are before downloading.
One practical tip: set the dictation language to Chinese explicitly rather than leaving it on automatic. Whisper decides the language from roughly the first thirty seconds and does not revisit that decision.