CMDF-TTS: Text-to-Speech Method with Limited Target Speaker Corpus

Abstract

While end-to-end Text-to-Speech (TTS) methods with limited target speaker corpus can generate high-quality speech, they often require a non-target speaker corpus (auxiliary corpus) which contains a substantial amount of <text, speech> pairs to train the model, significantly increasing training costs. In this work, we propose a fast and high-quality speech synthesis approach, requiring few target speaker recordings. Based on statistics, we analyzed the role of phonemes, function words, and utterance target domains in the corpus and proposed a Statistical-based Compression Auxiliary Corpus algorithm (SCAC). It significantly improves model training speed without a noticeable decrease in speech naturalness. Next, we use the compressed corpus to train the proposed non-autoregressive model CMDF-TTS, which uses a multi-level prosody modeling module to obtain more information and Denoising Diffusion Probabilistic Models (DDPMs) to generate mel-spectrograms. Besides, we fine-tune the model using the target speaker corpus to embed the speaker's characteristics into the model and Conditional Variational Auto-Encoder Generative Adversarial Networks(CVAE-GAN) to enhance further the synthesized speech's quality. Experimental results on multiple Mandarin and English corpora demonstrate that the CMDF-TTS model, enhanced by the SCAC algorithm, effectively balances training speed and synthesized speech quality. Overall, its performance surpasses that of state-of-the-art models.

Contents