Tsukasa 司 Speech: Engineering the Naturalness and Rich Expressiveness
tl;dr : I made a very cool japanese speech generation model.
if the demo didn't work and you just want to listen to some samples, take a look at this notebook. (ps. this belongs to a much earlier checkpoint, not representative of the model at its best.)
Try chatting with Aira, a mini-project I did by using various Tech, including Tsukasa. (maybe not very optimized, but hey, it works!)
Note: This model only supports the Japanese language; but you can feed it Romaji if you use the Gradio demo. (no longer, due to resource constraints, but the Tech is there.)
This is a speech generation network, aimed at maximizing the expressiveness and Controllability of the generated speech. at its core it uses StyleTTS 2's architecture with the following changes:
Incorporating mLSTM Layers instead of regular PyTorch LSTM layers, and increasing the capacity of the text and prosody encoder by using a higher number of parameters
Retrained PL-Bert, Pitch Extractor, Text Aligner from scratch
Whisper's Encoder instead of WavLM for the SLM
48khz Config
improved Performance on non-verbal sounds and cues. such as sigh, pauses, etc. and also very slightly on laughter (depends on the speaker)
a new way of sampling the Style Vectors.
Promptable Speech Synthesizing.
a Smart Phonemization algorithm that can handle Romaji inputs or a mixture of Japanese and Romaji.
Fixed DDP and BF16 Training (mostly!)
There are two checkpoints you can use. Tsukasa & Tsumugi 48khz (placeholder).
Tsukasa was trained on ~800 hours of studio grade, high quality data. sourced mainly from games and novels, part of it from a private dataset.
So the Japanese is going to be the "anime japanese" (it's different than what people usually speak in real-life.)
Special thanks to Yinghao Aaron Li, the Author of StyleTTS which this work is based on top of that. He is one of the most talented Engineers I've ever seen in this field.
Also Karesto and Raven(a.k.a hexgrad) for their help in debugging some of the scripts. wonderful people.
Why does it matter?
Recently, there's a big trend towards larger models, increasing the scale. We're going the opposite way, trying to see how far we can push the limits by utilizing existing tools.
Maybe, just maybe, scale is not necessarily the answer.
There's also a few things that's related to Japanese (but can have a wider impact on languages that face a similar issue like Arabic). such as how we can improve the intonations for this language. what can be done to accurately annotate a text that can have various spellings depending on the context, etc.
Third stage training (Kotodama, prompt encoding, etc.):
not planned right now, due to some constraints, but feel free to replicate.
some ideas for future
I can think of a few things that can be improved, not nessarily by me, treat it as some sorts of suggestions:
[o] changing the decoder (fregrad looks promising)
[o] retraining the Pitch Extractor using a different algorithm
[o] while the quality of non-speech sounds have been improved, it cannot generate an entirely non-speech output, perhaps because of the hard alignement.
[o] using the Style encoder as another modality in LLMs, since they have a detailed representation of the tone and expression of a speech (similar to Style-Talker).
Approximately 3 weeks of training, overall 3 months including the work spent on the data pipeline.
Roughly 66.6 kg of CO2eq. of Carbon emitted if we base it on Google Cloud. (I didn't use Google, but the cluster is located in US, please treat it as a very rough approximation.)
Kokoro - a very nice and light weight TTS, based on StyleTTS. supports Japanese and English. VoPho - a meta phonemizer to rule them all. it will automatically handle any languages with hand-picked high quality phonemizers.
litain's Moe Speech a very cool dataset you can use in case i couldn't release mine
@article{xlstm,
title={xLSTM: Extended Long Short-Term Memory},
author={Beck, Maximilian and P{\"o}ppel, Korbinian and Spanring, Markus and Auer, Andreas and Prudnikova, Oleksandra and Kopp, Michael and Klambauer, G{\"u}nter and Brandstetter, Johannes and Hochreiter, Sepp},
journal={arXiv preprint arXiv:2405.04517},
year={2024}
}
