Trained TTS Vitsmodel Restricted to 11 Seconds of Audio Generation

[shuvohishab]

By following this tutorial, I trained a TTS VitsModel.
While performing inference of the trained TTS model to convert text into speech, I observed a limitation where the system generates only up to 11 seconds of audio. So, when dealing with longer texts (e.g., equivalent to 90 seconds of audio), it just produces only first 11 seconds of equivalent audio.
On the other hand, when the text is short and equivalent to 0 to 11 seconds of audio, the system generates audio of the exact length.

Is there a workaround available to generate the complete audio for longer texts?

Note:

• I'm using the latest codebase to implement the TTS model.
• The audio data I am working with varies in length from 0.1 seconds to 63 seconds.

[shuvohishab]

@VahidooX could you kindly review my inquiry? Thank you.

[titu1994]

@XuesongYang

[XuesongYang]

thanks. @treacker. Could you pls have a look?

[XuesongYang]

found the same reported issue as well: #6998

[treacker]

Sorry for late responce. The problem is in max_len parameter, which is set by default to 1000 and can't be changed from convert_text_to_waveform function. The workaround is either to use forward function which is called inside convert_text_to_waveform or add option to change max_len. @XuesongYang choose please what is preferred solution

[PhilipAmadasun]

@shuvohishab Doesn't that use cpu? And also where is the audio gonna get stored? I want a .wav file out of this.

[shuvohishab]

Change cpu to cuda if you want to infer in GPU.
At first generate the audio, then you can save the audio with audio_pred .

[PhilipAmadasun]

@shuvohishab

audio_pred = model.convert_text_to_waveform(tokens=tokens).cuda().detach().numpy()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: can't convert cuda:0 device type tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.

[shuvohishab]

Kindly follow the instruction provided in error log.

[PhilipAmadasun]

@shuvohishab Never mind, forget my last message. I did have another question about the NeMo installation. When I changed the max lengths, did I need to reinstall nemo all over again? I pip installed the previous time.