Adds multispeaker support to inference_onnx with VITS, makes onnx inference faster

[Iamgoofball]

Adjusts the provider choice for CUDAExecutionProvider based on microsoft/onnxruntime#12880 (comment) and #2563 (comment)

Fixed onnx inference to support multispeaker.

Benched it against a personal multispeaker model with the following setup:

time_list = []
text_inputs_list = []
num_speakers = range(265)
print("Beginning 500 32 buffalo random-speaker inference test with text inputs calc'd each time")
for i in range(500):
    starttime = time.time()
    text = "Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo Buffalo."
    text_inputs = np.asarray(
        vits.tokenizer.text_to_ids(text, language="en"),
        dtype=np.int64,
    )[None, :]
    text_inputs_list.append(time.time()-starttime)
    starttime = time.time()
    audio = vits.inference_onnx(text_inputs, speaker_id=random.choice(num_speakers))
    time_list.append(time.time()-starttime)
    save_wav(wav=audio[0], path="./buffalo_samples/coqui_vits _" + str(i) + ".wav", sample_rate=config.audio.sample_rate)
print("Mean text_inputs time: " + str(statistics.mean(text_inputs_list)))
print("Mean inference time: " + str(statistics.mean(time_list)))
print("Total inference time: " + str(sum(time_list))))

We were able to generate 2 hours and 15 minutes of roughly 15-20 second clips within 32 seconds on an RTX 4080 using this setup.

Beginning 500 32 buffalo random-speaker inference test with text inputs calc'd each time
Mean text_inputs time: 0.03477813768386841
Mean inference time: 0.06550791311264038
Total inference time: 32.75395655632019

[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.}

[erogol]

thanks for the PR @Iamgoofball

Do you plan to sign the CLA?

[Iamgoofball]

I don't like putting my real name on things on this account, sorry. If you wanna mirror the PR though go for it, it's like 2 lines anyways and one of them's just borrowed from the comment linked in the OP about onnx params, I'll close this one when you do.

[jdlproenca]

Thank you @Iamgoofball !
Although it might be a rare use-case, I have a multi-speaker model with only 1 speaker, and I had to make one extra change in export_onnx from if self.num_speakers > 1: to if self.num_speakers > 0: here:

TTS/TTS/tts/models/vits.py

Line 1848 in 4bebd23

if self.num_speakers > 1:

Does it make sense in general? I see more if self.num_speakers > 0: in vits.py

[erogol]

Closing this for. #2743

[Nanayeb34]

what version of TTS are you using? @Iamgoofball