Mixer-TTS: non-autoregressive, fast and compact text-to-speech model conditioned on language model embeddings

TL;DR

Mixer-TTS introduces a non-autoregressive, fast, and compact speech synthesis model based on MLP-Mixer architecture, incorporating language model embeddings for improved performance and efficiency.

Contribution

The paper presents Mixer-TTS, a novel non-autoregressive TTS model using MLP-Mixer architecture and language model embeddings, achieving high-quality speech with fewer parameters and faster synthesis.

Findings

Achieves MOS of 4.05 and 4.11 for basic and extended versions.

Faster speech synthesis compared to similar quality models.

Uses fewer parameters, enabling compact deployment.

Abstract

This paper describes Mixer-TTS, a non-autoregressive model for mel-spectrogram generation. The model is based on the MLP-Mixer architecture adapted for speech synthesis. The basic Mixer-TTS contains pitch and duration predictors, with the latter being trained with an unsupervised TTS alignment framework. Alongside the basic model, we propose the extended version which additionally uses token embeddings from a pre-trained language model. Basic Mixer-TTS and its extended version achieve a mean opinion score (MOS) of 4.05 and 4.11, respectively, compared to a MOS of 4.27 of original LJSpeech samples. Both versions have a small number of parameters and enable much faster speech synthesis compared to the models with similar quality.