How Alice, intelligent assistant, learned to understand people with speech impairments
summary
Yandex upgrades Alice with a July 2024 model based on YandexGPT and real-world retraining on speech-impaired voices, using Yandex Tasks to collect about 1 million recordings (~900 hours) from 300+ participants classified by speech therapists. Training yields a 20% reduction in the accuracy gap versus typical speech (and an earlier 14% WER gain on a pilot), with recognition on par with or better than human transcribers for severe pronunciation impairments.
AI-written summary · about
chapters (12)
- 0:00Inclusive smart assistant intro
- 3:45Voice assistants accessibility
- 5:09Speech impairments user needs
- 7:18Project goals data collection
- 10:32Participant recruitment data
- 13:07Model performance evaluation
- 13:49Inclusive product recommendations
- 17:02Q&A project timeline
- 18:27Sensitive data challenges
- 19:55User feedback post-production
- 20:59Future global expansion
- 23:10Closing remarks Q&A
related talks
35mInside the New Generation of Alisa
50mAI as an A11y Ally
32mHow Alexa keeps getting smarter
49mThe impact of AI on people with disabilities
01h 06mAI as a Force for Innovation
46mArtificial intelligence to the rescue of accessibility
28mAI’s Next Great Challenge: How to create AI that truly understands us?
19mLLM-Enhanced Multimodal AI: Revolutionizing Audio/Video Interaction