Adds support for precomputing conditioning latents to xtts for repeated inference on the same reference wavs for significant performance gains.

[Iamgoofball]

Accidentally includes #2951 because I was lazy.

To use:

xtts = TTS("tts_models/multilingual/multi-dataset/xtts_v1", gpu=True)
xtts.synthesizer.tts_model.precompute_conditioning_latents("./path_to_wavs")

...

xtts.tts_to_file(
	text, 
	file_path="./path_to_output.wav", 
	speaker_wav="./xtts_ref_wavs/ref_speaker.wav", 
	language="en"
	precomputed_latents = True)

During benchmarking of the inference stack I found that like 2/3rds of the inference time was precomputation of the GPT latents. This is bad when you're working with fixed reference audio.

Without precomputed latents:

With precomputed latents:

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[skshadan]

pls provide full code.

[Iamgoofball]

this is the full code for this feature

[skshadan]

I used it still my inference time is same:(

[Iamgoofball]

i recommend doing some local benchmarking

[skshadan]

locally it is taking too much time:((

[skshadan]

[PranjalyaDS]

I have noticed that the latent computation is the longest during the first run, and the subsequent latent computations are taking time in a miniscule format, even if different reference audio files are used. I am not sure of the reason, the most obvious could be that the GPT and Diffusion latent generators are not lazy loaded, most probably?

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions. You might also look our discussion channels.