Inside the New Generation of Alisa
summary
Presentation describes the engineering of the new-generation Alice voice assistant: using techniques to reduce response latency so replies start in far less than ten seconds, lowering per-request compute cost to avoid spending thousands of GPU-hours, and integrating the new system without breaking existing stable services. The talk covers practical solutions and trade-offs the team used to achieve speed, cost-efficiency, and reliability.
AI-written summary · about
chapters (26)
- 0:00Introduction to Alisa
- 1:29How NLU and logic work
- 4:17Improving model responses
- 5:22Model optimization
- 7:13Reducing unnecessary queries
- 8:43Load reduction
- 9:51Speculative decoding
- 11:20Partial responses
- 13:28GPU transition
- 15:13Emotional responses
- 16:56Final architecture
- 17:20Q&A on censorship
- 18:23Child voices and parental control
- 19:50Bypassing censorship
- 20:11Alice wake word
- 21:33Response generation
- 21:53Foreign languages
- 22:59Emotional responses
- 24:05Quick response system
- 26:12Consecutive commands
- 28:01Memory overflow
- 29:27Session context
- 30:52Resources and optimization
- 32:16Integration plans
- 33:20Resource consumption trends
- 34:23Q&A and closing
related talks
53mAlisa, Don’t Lag!
24mHow Alice, intelligent assistant, learned to understand people with speech impairments
19mReal-Time AI: Low Latency Solutions for Interactive Applications
26mInterview
18mAlexa's next AI trick: Disappearing
26mEnabling the next generation in voice technology
17mServing Voice AI at Scale
12mBuilding Voice AI for People Who Don’t Trust AI