Deep learning-based speaker-adaptive postfiltering with limited adaptation data for embedded text-to-speech synthesis systems

Abstract

End-to-end (e2e) speech synthesis systems have become popular with the recent introduction of text-to-spectrogram conversion systems, such as Tacotron, that use encoder–decoder-based neural architectures. Even though those sequence-to-sequence systems can produce mel-spectrograms from the letters without a text processing frontend, they require substantial amounts of well-manipulated, labeled audio data that have high SNR and minimum amounts of artifacts. These data requirements make it difficult to build end-to-end systems from scratch, especially for low-resource languages. Moreover, most of the e2e systems are not designed for devices with tiny memory and CPU resources. Here, we investigate using a traditional deep neural network (DNN) for acoustic modeling together with a postfilter that improves the speech features produced by the network. The proposed architectures were trained with the relatively noisy, multi-speaker, Wall Street Journal (WSJ) database and tested with unseen speakers. The thin postfilter layer was adapted with minimal data to the target speaker for testing. We investigated several postfilter architectures and compared them with both objective and subjective tests. Fully-connected and transformer-based architectures performed the best in subjective tests. The novel adversarial transformer-based architecture with adaptive discriminator loss performed the best in the objective tests. Moreover, it was faster than the other architectures both in training and inference. Thus, our proposed lightweight transformer-based postfilter architecture significantly improved speech quality and efficiently adapted to new speakers with few shots of data and a hundred training iterations, making it computationally efficient and suitable for scalability.