Spoken Chinese languages

[kruton]

The list of languages mentions Chinese which is imprecise when dealing with spoken languages which might share similar writing systems. Is it trained on spoken Mandarin, Cantonese, Hakka, something else, some of these, all of these, etc?

[jongwook]

When using zh (Chinese) as the language label, our medium and large models showed ~8% WER and a BLEU score of ~13 on the yue_hant_hk split of the Fleurs dataset, suggesting that it works quite decently on Cantonese and possibly on other dialects as well. We'd be interested in knowing how usable it is for the other Sinitic languages.

[ronyfadel]

How can you get transcriptions for Taiwan (traditional Chinese characters) v/s China (simplified Chinese characters)?

[glangford]

fyi

• How do I choose whether the subtitles are spoken language or written language? #366 (comment)

[miacias]

Cantonese is not a dialect; it is a distinct language from Mandarin. Both, among others, make up the 'Chinese' or Sinitic branch of the Sino-Tibetan language family. @kruton asked a valid question about whether the label marked as zh Chinese will make a distinction between these separate languages.

[ryanheise]

The large-v3 model was just announced (#1762 ) which introduces a separate language code for Cantonese. You can use yue or Cantonese.

Models prior to large-v3 (as mentioned above) are capable of transcribing both Mandarin and Cantonese (and possibly others), even though there was just a single zh label. Although with v3, the accuracy for Cantonese should be greater under the yue label than it was in v2 under the zh label.

[miacias]

Thank you so much for this info, this is exactly what I was hoping for!