Outputs too many <unk> symbols with Mandarin Chinese (cmn & cmn_Hant) and Cantonese (yue)

[tanshuai]

For example "Oh, Peter." translated to "\<unk>,彼得.", "Oh, my god" translated to "\<unk>,我的上帝"

Almost all of the "Oh" are translated into \<unk>, making this project almost unusable for Chinese and Cantonese..

$ m4t_predict "Oh, Peter."  t2tt cmn --src_lang eng
2023-09-20 03:22:20,215 INFO -- m4t_scripts.predict.predict: Running inference on the GPU in torch.float16.
Using the cached checkpoint of the model 'seamlessM4T_large'. Set `force=True` to download again.
Using the cached tokenizer of the model 'seamlessM4T_large'. Set `force=True` to download again.
Using the cached checkpoint of the model 'vocoder_36langs'. Set `force=True` to download again.
2023-09-20 03:22:24,949 INFO -- m4t_scripts.predict.predict: Translated text in cmn: <unk>,彼得.

$ m4t_predict "Oh, Peter."  t2tt cmn_Hant --src_lang eng
2023-09-20 03:22:48,454 INFO -- m4t_scripts.predict.predict: Running inference on the GPU in torch.float16.
Using the cached checkpoint of the model 'seamlessM4T_large'. Set `force=True` to download again.
Using the cached tokenizer of the model 'seamlessM4T_large'. Set `force=True` to download again.
Using the cached checkpoint of the model 'vocoder_36langs'. Set `force=True` to download again.
2023-09-20 03:22:53,404 INFO -- m4t_scripts.predict.predict: Translated text in cmn_Hant: <unk>, 彼得.

$ m4t_predict "Oh, Peter."  t2tt yue --src_lang eng
2023-09-20 03:21:16,073 INFO -- m4t_scripts.predict.predict: Running inference on the GPU in torch.float16.
Using the cached checkpoint of the model 'seamlessM4T_large'. Set `force=True` to download again.
Using the cached tokenizer of the model 'seamlessM4T_large'. Set `force=True` to download again.
Using the cached checkpoint of the model 'vocoder_36langs'. Set `force=True` to download again.
2023-09-20 03:21:20,886 INFO -- m4t_scripts.predict.predict: Translated text in yue: <unk>,彼得.

152

64

[Qubitium]

Confirmed also on my end with SeamlessM4TLarge model.

English input:

A witch can fly with a broom.

Chinese output:

一个女巫可以用扫<unk>飞.

Seamless t2t is based on nllb but nllb does not have this issue.

[Qubitium]

@cndn you pushed a readme update at https://github.com/facebookresearch/seamless_communication/commit/df2816adf3574016ffa99eb947ec3bff23310413 but the diff changes is related to audio alignment while this issue is text2text translation. Can you provide an example for we can fix this for t2t? Thanks.

[aliencaocao]

Same issue here on v2 model T2TT...How can the model be predicting a special token in the first place? This does not happen on the HF space demo, and I am using the same code as the demo... It only happens for some instances of Oh for me

[asulada]

me too

[skywindy]

how can i fix this problem?

[Qubitium]

@skywindy I believe the tokenizer mapping in the opensourced repo is completely wrong. Thus creating this UNK issues. Either that or they trained with a broken tokenizer. This is no way for us to fix this without knowing if model or tokenizer is broken.