zoom_in Click to enlarge

WaveNet

▲ ▼

description WaveNet Overview

WaveNet is a deep neural network architecture for generating raw audio waveforms, introduced by researchers at DeepMind in a paper published in September 2016. It uses a dilated causal convolutional network to model audio signals directly at the sample level, producing more natural-sounding speech than prior concatenative and parametric text-to-speech systems. The approach also enabled generation of music and other audio. WaveNet's architecture influenced subsequent neural audio synthesis systems and was later adapted for use in Google Assistant voice synthesis.

help WaveNet FAQ

What makes WaveNet different from previous text-to-speech systems?

Introduced by DeepMind in 2016, WaveNet is a deep neural network that generates raw audio waveforms directly at the sample level. Previous systems relied on concatenative or parametric methods, which often produced robotic-sounding speech, whereas WaveNet produces highly natural, human-like audio.

What underlying architecture does WaveNet use to process audio?

It utilizes a dilated causal convolutional network, allowing the model to have a massive receptive field without needing an excessive number of parameters. This specific architecture enables WaveNet to understand long-range temporal dependencies in the raw audio signal.

Was WaveNet ever integrated into consumer-facing applications?

Yes, a significantly sped-up version of the WaveNet architecture was integrated into Google Assistant to power its speech synthesis. This allowed Google's virtual assistant to speak with much more natural-sounding, human-like intonation.

How does WaveNet handle music generation?

In addition to speech, WaveNet can generate instrumental music by training on datasets of classical piano and symphonic recordings. By predicting the next audio sample, it can compose original musical passages that mimic the stylistic patterns of its training data.

swap_horiz

Looking for WaveNet alternatives? Compare top competitors ranked & scored

arrow_forward