Google’s new text-to-speech model, Tacotron, shows a lot of promise. We spent last night listening to the audio samples generated by it. In particular, the examples where the model performed punctuation inflections were really impressive. Tacotron’s architecture combines seq2seq methods with an attention mechanism. It’s trained from scratch on text-audio pairs without any specific feature engineering. Unfortunately, they use an internal dataset so you can’t go out an replicate their work today.
Text-to-Speech from Google
Post · Mar 30, 2017 16:54 · Share on Twitter
A text-to-speech synthesis system typically consists of multiple stages, such as a text analysis frontend, an acoustic model and an audio synthesis module. Building these components often requires extensive domain expertise and may contain brittle design choices. In this paper, we present Tacotron, an end-to-end generative text-to-speech model that synthesizes speech directly from characters. Given <text, audio> pairs, the model can be trained completely from scratch with random initialization. We present several key techniques to make the sequence-to-sequence framework perform well for this challenging task. Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English, outperforming a production parametric system in terms of naturalness. In addition, since Tacotron generates speech at the frame level, it's substantially faster than sample-level autoregressive methods.