ECAPA-TDNN for Multi-speaker Text-to-speech Synthesis

Abstract

In recent years, neural network based methods for multispeaker text-to-speech synthesis (TTS) have made significant progress. However, the current speaker encoder models used in these methods still cannot capture enough speaker information. In this paper, we focus on accurate speaker encoder modeling and propose an end-to-end method that can generate highquality speech and better similarity for both seen and unseen speakers. The proposed architecture consists of three separately trained components: a speaker encoder based on the state-ofthe-art ECAPA-TDNN model which is derived from speaker verification task, a FastSpeech2 based synthesizer, and a HiFi-GAN vocoder. The comparison among different speaker encoder models shows our proposed method can achieve better naturalness and similarity. To efficiently evaluate our synthesized speech, we are the first to adopt deep learning based automatic MOS evaluation methods to assess our results, and these methods show great potential in automatic speech quality assessment.

Simple and Effective Unsupervised Speech Synthesis

“…Speaker modeling on LibriTTS: Following prior work [26, 2 https://github.com/pytorch/fairseq/tree/main/ examples/wav2vec/unsupervised 3 https://github.com/espnet/espnet/tree/master/egs2/ 27, 28], a pre-trained speaker verification model ECAPA-TDNN [29] from SpeechBrain [30] is used as the speaker encoder to generate speaker embeddings. The Transformer-TTS is conditioned on the speaker embedding and Global Style Tokens (GST) [31] for multi-speaker modeling [27,28]. We found that fine-tuning the TTS on each target speaker for 10 epochs significantly improves the synthesis quality.…”

A Novel End-to-End Turkish Text-to-Speech (TTS) System via Deep Learning

“…The main objective of developing TTS systems with this method is to capture the speaker characteristics by extracting text-independent speaker embeddings from the target speaker's voice. Commonly used speaker encoders in TTS systems include d-vector and x-vector [26]. In order for these methods to be adapted to Turkish speech synthesis systems, a large-sized speech-text corpus or pre-trained acoustic models are needed [26].…”

“…Commonly used speaker encoders in TTS systems include d-vector and x-vector [26]. In order for these methods to be adapted to Turkish speech synthesis systems, a large-sized speech-text corpus or pre-trained acoustic models are needed [26]. Currently, the lack of a Turkish corpus or a previously trained acoustic model available makes it difficult to process Turkish in the field of TTS.…”