data.bythebay.io: Shubho Sengupta,
summary
Claims a practical end-to-end deep speech system that replaces complex ASR pipelines, trained on 12,000+ hours (synthetic to ~100,000) and delivering state‑of‑the‑art Mandarin recognition that beats humans on short queries while struggling on heavily accented English. Built a custom HPC stack: distributed linear algebra (Majel), MPI‑based sync communication, GPU‑only training (32–64 GPUs typical), a 400× faster warp‑CTC GPU implementation, custom memory allocator and InfiniBand fat‑tree cluster. Models run as hundreds of millions of parameters; deployment uses smart batch dispatch to meet latency constraints.
AI-written summary · about
chapters (8)
- 0:00Introduction and motivation
- 2:22Mobile adoption in China
- 6:55End-to-end speech recognition
- 13:55High-performance training system
- 21:25Multilingual training success
- 27:43Model deployment challenges
- 32:00Future research directions
- 40:40Q&A on accents and lip reading
related talks
27mThe Story of End to End Models in Deep Learning
16mThe End of Awkward AI Transcriptions
21mMultimodal Open Source at Kyutai, From Online Demos to On-Device
22mBuilding next-gen speech synthesis for Bookmate audiobook service
35mTricks Learned from Scaling WhisperSpeech Models to 80k+ Hours of Speech
21mDeveloping Machine Learning Based Speech Enhancement Models for Teams and Skype
47mThe Digital Hearth: An In-the-Wild Automated Group Speech to Text System
08mQ n A with Dr Ananth Sankar