Spaces:

Duplicated from  TTS-AGI/TTS-Arena

Pendrokar

/

TTS-Spaces-Arena

like 114

Running

@hexgrad seems to be just Kokoro.
hexgrad/kokoro: [WARN] Unable to resample audio

[edit] meaning no one can vote on Kokoro samples as in this Arena it is required to listen to both samples.

Fails before even trying to normalize the sample, code part:
https://huggingface.co/spaces/Pendrokar/TTS-Spaces-Arena/blob/main/app.py#L1196-L1213

This does not return a sample for me.

import os
from gradio_client import Client, file

client = Client("hexgrad/kokoro", hf_token=os.getenv('HF_TOKEN'))

result = client.predict(
        text="The juice of lemons makes fine punch.",
        voice="af",
        ps=None,
        speed=1,
        trim=3000,
        use_gpu=False,
        sk=None,
        api_name="/generate"
)
print(result)

(None, '')

[EDIT] Wait holdup, I forgot that only the hardcoded phrases work. Changed the text and it works.

New bug without any changes made to Gradio. A generated sample is not returned. Still gets saved as cached sample.

I can confirm seeing this bug, but didn't know which model it was because I can't even vote.

Kokoro should be fixed and has been tested. @Pendrokar @ecyht2

🏆 2024-12-02 18:53:01.058352 ['af'] 60 False 88340

Unfortunately, it seems that empty audios have already been cached. To avoid this in the future, I have changed it to an assert which raises an error instead of returning (and caching) empty audio.

Seems OK now.