Architecture
Is this Tacotron-like?
There is a niche for small single language TTS models. Pity is all of them are English only. If I make a Bulgarian dataset, do you have a guide/scripts/tutorial on how to train a model with similar quality to yours for that other language?
If espeak supports it, then it if fine. You can literally use any architecture. How many hours do you have? :)
Prepared and cleaned, none so far, but this small model seemed like good starting point to train something.
Its not really Tacotron-like, its closer to a tiny FastSpeech-style two-stage stack:
bascially: text/phonemes -> compact non-autoregressive acoustic model -> mel spectrogram -> small HiFi-GAN-style vocoder -> waveform
So there is no autoregressive decoder like Tacotron. The acoustic model predicts duration/prosody-ish features and generates mels, then the vocoder turns those mels into 24 kHz audio.
Prepared and cleaned, none so far, but this small model seemed like good starting point to train something.
I think maybe try waiting a bit, I'm already working on Inflect-Nano-v2! Not completely decided yet, but it will likely be slightly larger at about 6-8m parameters instead of 4.6m, I think its a better sweet spot. And a lot of the files and stuff were honestly just really unorganized with this first version. I didn't really expect it to blow up as much as it did. I will make sure the v2 is much easier for people to finetune and use though, and it'll likely fix a lot of the problems with v1, especially with pronunciation and artifacts, maybe at the cost of some expressiveness.