Generative Model-Based Text-to-Speech Synthesis

Explore the latest advancements in generative model-based approaches for speech synthesis in this 38-minute conference talk by Heiga Zen from Google. Gain insights into the significant improvements in synthesized speech naturalness, learn about the probabilistic formulation of text-to-speech systems, and discover various acoustic models including HMM-based, FFNN-based, and NN-based generative models. Delve into the architecture of WaveNet, a groundbreaking generative model for raw audio, and understand its advantages over conventional audio generative models. Examine the potential future directions in text-to-speech synthesis and its applications beyond traditional boundaries.

Intro
Outline
Text-to-speech as sequence-to-sequence mapping
Speech production process
Typical flow of TTS system
Speech synthesis approaches
Probabilistic formulation of TTS
Approximation (2)
Representation - Linguistic features
Representation - Acoustic features
Representation - Mapping
HMM-based generative acoustic model for TTS
Alternative acoustic model
FFNN-based acoustic model for TTS [6]
NN-based generative acoustic model for TTS
NN-based generative model for TTS
Learned features
WaveNet: A generative model for raw audio
WaveNet - Causal dilated convolution
WaveNet - Architecture
Softmax
WaveNet vs conventional audio generative models
Relax approximation
Generative model-based text-to-speech synthesis
Beyond text-to-speech synthesis
Beyond generative TTS