Load testing speech synthesis and recognition at SberDevices
summary
Describes how SberDevices performs load testing for its speech synthesis and recognition pipelines to handle billions of seconds of audio with guaranteed throughput. Covers the engineering challenges, metrics and test setups used to simulate large-scale continuous processing, and tools/approaches for ensuring reliable, high-performance speech I/O at production scale.
AI-written summary · about
chapters (14)
- 0:00Load testing intro
- 1:51Purpose of testing
- 3:36Simple audio testing
- 5:45Real-time audio emulation
- 8:42End-of-phrase detection
- 11:10Accumulated delay in chunks
- 14:19Server saturation limits
- 16:27Speech synthesis testing
- 18:37Real-time synthesis delays
- 21:07Measuring synthesis delay
- 23:40Buffer starvation in synthesis
- 24:43Caching in synthesis
- 26:32Methodology summary
- 27:35Q&A
related talks
42mdata.bythebay.io: Shubho Sengupta,
35mTricks Learned from Scaling WhisperSpeech Models to 80k+ Hours of Speech
46mIntroduction to Speech Recognition Systems Through a Tester’s Eyes
24mTowards Large-scale Speech Analytics Systems
56mSpeech Recognition: How to Build Speech-to-Text Yourself
26mEnabling the next generation in voice technology
17mServing Voice AI at Scale
25mEngineering voice agents: Latency, quality, and scale