Research on Speech Synthesis Based on Mixture Alignment Mechanism

In recent years, deep learning-based speech synthesis has attracted a lot of attention from the machine learning and communities. this paper, we propose Mixture-TTS, non-autoregressive model based on mixture alignment mechanism. Mixture-TTS aims to optimize information between text sequences mel-spectrogram. uses linguistic encoder soft phoneme-level hard word-level approaches, which explicitly extract semantic information, introduce pitch energy predictors optimally predict rhythmic audio. Specifically, introduces post-net five-layer 1D convolution network reconfiguration capability We connect output decoder through residual network. The mel-spectrogram is converted into final audio by HiFi-GAN vocoder. evaluate performance AISHELL3 LJSpeech datasets. Experimental results show that somewhat better in mel-spectrogram, able achieve high-quality ablation studies demonstrate structure effective.