Microsoft · 2023 · codec LM

VALL-E

TTS as language modelling over neural codec tokens (EnCodec-class), trained on 60k hours. Three seconds of prompt audio become a speaker. The clip is the official demo introducing VALL-E in a Bill Gates prompt voice — a vendor/research demo, not our bench script.

Known for

Zero-shot cloning as in-context learning.

Vendor / paper demo

Bill Gates prompt (zero-shot) · VALL-E official demo page, audios/bgreview/billg_valle.wav.

Official VALL-E demo: ‘Introduce VALL-E with Bill Gates's voice.’ Zero-shot clone from a short prompt. Not the shared bench script.

Clip source

Approach: Codec LM
Year / era: 2023
License class: research
Cloning: Yes
Streaming: No
Emotion controls: Yes
Languages (tags): en
Official: Source