WaveNet Turns 10: Authors Reflect on a Decade of Speech Synthesis Paradigm Shifts
WaveNet marks its 10th anniversary. Published by Google DeepMind in 2016, it generated raw speech waveforms autoregressively with long context, displacing the then-dominant concatenative and statistical parametric synthesis paradigms with striking quality. Co-author Heiga Zen and DeepMind researcher Sander Dieleman both published retrospectives tracing the field's evolution from waveform generation to end-to-end systems and multimodal LLMs.
Confirmed
- WaveNet was released by DeepMind in 2016; Heiga Zen collaborated on it, and its direct raw-waveform generation upended concatenative and statistical parametric synthesis.
- Sander Dieleman recalled that the autoregressive, long-context design predated Transformers simply because Transformers had not yet been invented (introduced in 2017).
- Dieleman noted that "audio generation as language modeling" is now the industry-standard paradigm, and the dilated convolutions from the paper remain widely used in audio codecs and vocoders.
- Researcher @santty128 cited Dieleman's retrospective, remarking that TTS paradigms have completely shifted as end-to-end methods broke multiple inductive biases of speech synthesis.
- Heiga Zen's retrospective outlined the decade's arc: WaveNet's waveform generation, the end-to-end era, and today's multimodal LLM era of speech synthesis.
Why it matters
- WaveNet is a milestone in speech and audio generation; the autoregressive paradigm it established has shaped a decade of audio foundation models, codecs, and vocoder design.
- First-hand retrospectives from the original authors offer a rare perspective on how speech synthesis migrated from waveform modeling to multimodal LLMs.
2026-09-09 ~ 2026-09-10 · 5 related posts
Primary sources
- Ten years of WaveNet: Google researcher traces speech synthesis from raw waveforms to multimodal LLMs — heiga_zen · 2026-09-09
- WaveNet turns 10: long-context autoregression a year before Transformers existed — teropa · 2026-09-09
- [source] WaveNet turns 10: audio-as-language-modeling is now the standard, dilated convs still everywhere — heiga_zen · 2026-09-09