Speech to text

/

POST /v1/audio/transcriptions takes multipart/form-data, not JSON: send model as a form field and the recording as a file part named file. The reply is {"text": "…"}. POST /v1/audio/translations is the same shape and returns English. Unlike the video endpoints this is synchronous, so a long recording holds the connection open until it finishes.

POST /v1/audio/transcriptions 
curl https://api.chinaapi.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $CHINAAPI_KEY" \
  -F model=stepaudio-2.5-asr \
  -F file=@speech.mp3

# {"text":" China api text to speech verification."}
Lowest cost

stepaudio-2.5-asr

With stepaudio-2-asr-pro. stepaudio-2.5-asr is by a wide margin the cheapest transcription model on the endpoint; the pro variant costs roughly fifteen times as much and is worth it only when accuracy on difficult audio matters.

Stream variants

stepaudio-2.5-asr-stream

This model is built for streaming recognition, but this endpoint does not stream its output: upload a complete wav (16-bit PCM) or ogg recording and the whole transcript comes back in one response, as with the other models. It costs more per minute than its non-streaming counterpart.

Multilingual and dialects

qwen3-asr-flash

qwen3-asr-flash supports multilingual transcription. mimo-v2.5-asr is tuned for Chinese dialects.

Tip Billing counts the length of the audio you upload, at a thousand tokens to the minute, and the duration is rounded up to a whole second before that conversion — so a five second clip bills as 83 tokens and anything under a second still bills as if it were one. The transcript itself is free. Cost therefore scales with recording length and not with how much speech the recording contains, which makes trimming silence worthwhile.