FoundationTTS: Text-to-Speech for ASR Customization with Generative Language Model

Neural text to speech (TTS) generally consists of cascaded architecture with separately optimized acoustic model and vocoder or end-to-end architecture with continuous mel-spectrograms or self-extracted speech frames as the intermediate representations to bridge acoustic model and vocoder in joint training, which suffers from two limitations: 1) the continuous acoustic frames are hard to predict with phoneme only, acoustic information are also needed like duration or pitch to solve the one-to-many problem, which is not easy to scale on large scale and noise dataset; 2) diverse speech output is not straightforward with continuous speech features and complex VAE or flow based models are often needed. In this paper, we propose FoundationTTS, a new speech synthesis system with discrete speech tokens extraction from neural audio codec and a large language modelling based acoustic model for simultaneously optimizing linguistic and acoustic tokens. Specifically, 1) we propose a hierarchical codec network based on vector-quantized auto-encoders with adversarial training (VQ-GAN) to first extract continuous frame-level speech representations with fine-grained codec, and the coarse-grained codec reconstructs the continuous speech frame with fewer quantizers; 2) we jointly optimize speech token, linguistic tokens, speaker token together with a large language model and autoregressively predict the discrete speech tokens. Experiments show that FoundationTTS achieves a MOS gain of +0.14 compared to the baseline system. In ASR customization tasks, our method achieves 7.09\% and 10.35\% WERR respectively over two strong customized ASR baselines.