Differentiable Duration Refinement Using Internal Division for Non-Autoregressive Text-to-Speech

Most non-autoregressive text-to-speech (TTS) models acquire target phoneme duration (target duration) from internal or external aligners. They transform the speech-phoneme alignment produced by the aligner into the target duration. Since this transformation is not differentiable, the gradient of the loss function that maximizes the TTS model's likelihood of speech (e.g., mel spectrogram or waveform) cannot be propagated to the target duration. In other words, the target duration is produced regardless of the TTS model's likelihood of speech. Hence, we introduce a differentiable duration refinement that produces a learnable target duration for maximizing the likelihood of speech. The proposed method uses an internal division to locate the phoneme boundary, which is determined to improve the performance of the TTS model. Additionally, we propose a duration distribution loss to enhance the performance of the duration predictor. Our baseline model is JETS, a representative end-to-end TTS model, and we apply the proposed methods to the baseline model. Experimental results show that the proposed method outperforms the baseline model in terms of subjective naturalness and character error rate.