Advance diversity & inclusion in voice AI with speech disentanglement
summary
Amazon describes concrete ML work: a neural TTS pipeline with encoder‑attention‑decoder and vocoder that adds speaker ID, language ID, accent ID to produce a single voice across languages (a polyglot model), uses voice‑conversion and flow‑based decoders to synthesize accents and interpolate speakers, and trains a DNN converter for whisper mode. Deployments include a masculine voice in five locales, Reading Sidekick, preferred speaking‑rate control, and multilingual mode covering 15 language pairs; goal is to shrink measured gaps like a cited 16 percentage point ASR accuracy disparity.
AI-written summary · about
chapters (13)
- 0:00Introduction to DEI in voice AI
- 6:55Addressing gender diversity gaps
- 9:07Supporting multilingual households
- 11:38Inclusivity for disabilities and kids
- 14:29Future vision for inclusive voice AI
- 17:26Text-to-speech technology overview
- 24:17Speech disentanglement explained
- 27:55Polyglot multilingual TTS model
- 33:41Creating accented voices
- 36:12Flow-based decoder for voice control
- 38:47Reading Sidekick for children
- 41:56Accessibility features like whisper mode
- 46:31Summary of DEI advancements
related talks
52mGoing Full Duplex - Building Apps for Speech
28mWhat’s Needed in Tech to bring Voice to the Future
17mBuilding Effective Voice Agents
01h 02m[Full Workshop] Building Conversational AI Agents
43mBuild Voice-Enabled Experiences
10mBuilding Conversational AI Agents at Scale
44mSmartly AI: Cross Platform Voice Application Creation
01h 06mAI as a Force for Innovation