VyvoTTS is an LLM-based text-to-speech (TTS) training framework developed by Vyvo-Labs and open-sourced on GitHub. Its core idea is to have a large language model generate interleaved audio codec tokens, then decode these tokens into audio waveforms, thereby completing the conversion from text to speech.
In terms of main capabilities, VyvoTTS supports two audio codecs, SNAC and Mimi. The framework provides four backends, and all backends share the same interface, so users only need to change the import to switch backends. The project also provides benchmarks based on Qwen3-1.7B and H100 PCIe.
In terms of usage, VyvoTTS covers the complete pipeline from data preparation to model training and generation. Dataset preparation supports standard dataset tokenization as well as large-scale Emilia dataset tokenization, and supports two Emilia data sources: ylacombe/emilia-subset (3.39 million English samples, parquet-based) and amphion/Emilia-Dataset (Emilia + Emilia-YODAS English, tar-based). Pretraining supports multi-GPU FSDP, and the model, tokenizer, codec type, dataset path (local or HuggingFace), and QA:TTS ratio scheduling (2:1 to 1:1) can be configured in vyvotts/configs/train/lfm2_5_pretrain.yaml. Single-speaker fine-tuning supports automatic tokenization, and the entire pipeline covers downloading, tokenizing with the codec, training, and generating test wav files, with a complete guide in FINETUNE.md. In addition, the framework also supports codebook-aware SFT, preference alignment, and online RL for quality alignment and WER post-training; examples are provided in the training guide, and the complete protocol and results are provided in the experiment report. Full training can be performed via accelerate.
Applicable scenarios include research and development work that requires building text-to-speech systems based on LLMs, especially scenarios involving large-scale dataset training, multi-GPU training, single-speaker fine-tuning, and quality alignment and WER optimization.
In terms of limitations, the current materials do not providecomplete information on charging, pricing, platform support scope, specific performance metrics, or detailed usage steps; some content requires reference to the configuration files and documentation in the project repository.
