Ten years of WaveNet: Google researcher traces speech synthesis from raw waveforms to multimodal LLMs
heiga_zen · x · 2026-09-09
Google speech researcher Heiga Zen marks the 10th anniversary of WaveNet with a retrospective on a decade of speech synthesis.
- WaveNet (2016): shattered the concatenative/statistical-parametric paradigm by directly generating raw audio waveforms
- End-to-end era: encoder-decoder models like Tacotron generated acoustic features from text, while text-to-waveform models like VITS and NaturalSpeech removed intermediate features entirely
- LLM convergence: AudioLM with SoundStream and VALL-E tokenized audio into the LLM framework, culminating in today's multimodal speech generation
A firsthand technical history from someone who lived it — equal parts tutorial and industry lore.
More from Research
- EdgeRock-4B holds KV cache at 72 MiB across 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- TRL adds Score Centering, a simple fix for RLHF training instability — lvwerra · 2026-10-27
- Latent-Foresight: world models that learn to predict their own future — abursuc · 2026-10-27
- UK AISI papers: no good safeguards yet against misuse of biological AI tools — davidmanheim · 2026-10-27
- CMU's XYBench: LLMs answer queries literally and reinforce user misconceptions — valentina__py · 2026-10-27
- arXiv advice from T. Dietterich: wrong primary category will almost surely get your paper on hold — tdietterich · 2026-10-27