R
Richard Kim
Creating an assistant with
transcriber.model: "nova-3"
and any Chinese language code returns 400. The documentation given at https://docs.vapi.ai/providers/transcriber/deepgram#supported-languages marks Chinese as supported on Nova 3, and the Python SDK's type union accepts "zh" for
language
.
"ru" and "es" results in 201 but "zh" or "zh-Hans" results in the following 400:
"400 Bad Request
assistant.transcriber.language must be one of the following values for nova-3 model:
en, en-US, en-AU, en-GB, en-IN, en-NZ, de, de-CH, nl, nl-BE, sv, sv-SE, da, da-DK, es,
es-419, fr, fr-CA, pt, pt-BR, pt-PT, it, tr, no, id, ar, be, bg, bn, bs, ca, cs, et, fi,
el, fa, he, hi, hr, hu, ja, kn, ko, ko-KR, lv, lt, mk, mr, ms, pl, ro, ru, sk, sl, sr, ta,
te, tl, uk, ur, vi, multi"
Deepgram's API accepts the combination of zh and nova-3. Posting Mandarin audio to https://api.deepgram.com/v1/listen?model=nova-3&language=zh returns valid transcripts, and they are noticeably more accurate than the same audio through nova-2.
Please look into this issue/temporarily update the documentation accordingly.