Difference between TextToSpeechDataset and TTSDataset

[bzp83]

What is the difference between those 2 and when to use each?

I see fastpitch configs using both but not sure which one to use... I did some training tests with the config using TextToSpeechDataset but it seems like it's not generating a good model...

[rlangman]

TTSDataset was the data loader FastPitch was originally implemented with several years ago. TextToSpeechDataset is a newer implementation that was intended to replace it, but which never made it past the experimental phase because we stopped working on older architectures like FastPitch. Both should produce the same result if configured correctly.

It "not generating a good model" could be a result of many things, depending on what specific behavior you are seeing from it.

[bzp83]

thank you @rlangman. The reason I'm using fastpitch now is because it's generating better audios than vits and we need to run inference on the CPU so we can't rely on bigger models because it would be slower than real time. I also fixed the "new" config that uses TextToSpeechDataset, it was just a mel config mismatch.

I'm going to evaluate RADTTS as well, are you able to tell me if it's better than fastpitch and if it's also considered older architecture?

Any suggestion on what else we could use for CPU inference? Thanks!

[rlangman]

FastPitch, RadTTS, and VITS I would call our previous-generation TTS models. But our current generation model (MagpieTTS) is too large for CPU.

For CPU, FastPitch is your best bet. RadTTS is less robust than FastPitch, but might sound a little more natural/expressive for some speakers.