# WaveNet: A Generative Model for Raw Audio

**Authors:** Aaron van den oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, Koray Kavukcuoglu  
**arXiv:** [1609.03499](https://arxiv.org/abs/1609.03499)  
**Date:** 2016-09-12  
**Summary page:** https://www.summarizepaper.com/en/paper/1609.03499v1/

## Key points

- "WaveNet: A Generative Model for Raw Audio" is a paper that introduces a deep neural network that generates raw audio waveforms.
- The model is fully probabilistic and autoregressive, meaning that the predictive distribution for each audio sample is conditioned on all previous ones.
- Despite its complexity, the model can be efficiently trained on data with tens of thousands of samples per second of audio.
- When applied to text-to-speech, WaveNet achieves state-of-the-art performance and human listeners rate it as significantly more natural sounding than other systems for both English and Mandarin.
- A single WaveNet can capture the characteristics of many different speakers with equal fidelity and switch between them by conditioning on the speaker identity.
- WaveNet can also generate novel and highly realistic musical fragments when trained to model music.
- It has potential applications in speech synthesis, music composition, and speech recognition.

## Summary

The paper "WaveNet: A Generative Model for Raw Audio" introduces a deep neural network that generates raw audio waveforms. The model is fully probabilistic and autoregressive, meaning that the predictive distribution for each audio sample is conditioned on all previous ones. Despite this complexity, the model can be efficiently trained on data with tens of thousands of samples per second of audio. When applied to text-to-speech, WaveNet achieves state-of-the-art performance, with human listeners rating it as significantly more natural sounding than the best parametric and concatenative systems for both English and Mandarin. Additionally, a single WaveNet can capture the characteristics of many different speakers with equal fidelity and switch between them by conditioning on the speaker identity. The authors also trained WaveNet to model music and found that it generates novel and often highly realistic musical fragments. Furthermore, they demonstrated that it can be employed as a discriminative model, returning promising results for phoneme recognition. Overall, WaveNet represents a significant advancement in generative models for audio processing and has potential applications in speech synthesis, music composition, and speech recognition.