Shinji Watanabe

Results 318 comments of Shinji Watanabe

Discrete token-based TTS implementation

Thanks for your great efforts, @ftshijt!

Error while using auxiliary CTC objective for multilingual ASR

Thanks for the report. @wanchichen, do you know what happens?

Diarization, missing example

We have an example here, but it is in an experimental stage. https://github.com/espnet/espnet/tree/master/egs2/librimix/enh_diar1 It is better to check the other tools, e.g., https://github.com/pyannote/pyannote-audio

load state_dict with map_location='cpu'

@pyf98 and @pengchengguo, can you double-check this (as it affects many parts). @Masao-Someki Thanks a lot! Can you add some comments to the code?

load state_dict with map_location='cpu'

Hi @pyf98 and @pengchengguo, This is a reminder. Can you check this PR? As I said, I'm careful about this PR since it affects many parts, and I want your...

load state_dict with map_location='cpu'

OK, thanks for the investigation. Should we document `torch.cuda.empty_cache()`?

Inefficient GPU memory allocation for S2T inference?

Thanks, @Masao-Someki, for the report! @pyf98, can you confirm it? @Masao-Someki, will it be applied to the other inference or only s2t?

Inefficient GPU memory allocation for S2T inference?

Thanks for the report. Can you make a PR? It seems that we should fix it.

Having output discrepancy in stage 7 of multi-task enh model

Thanks for the report! @simpleoier, could you double-check this? I also feel that this sounds strange.

RuntimeError: Internal: src/trainer_interface.cc(580) [(static_cast<int>(required_chars_.size() + meta_pieces_.size())) <= (trainer_spec_.vocab_size())] Vocabulary size is smaller than required_chars. 30 vs 65. Increase vocab_size or decrease character_coverage with --character_coverage option.

Please increase the vocabulary size by using `--nbpe` option