The End of Awkward AI Transcriptions
summary
Claims concrete results and usable tooling: 6 top-ten models on the Hugging Face ASR leaderboard, a Parakeet2 family for streaming and a Rea Canary family for high-accuracy multitask models, and a fast conformer backbone with 80 ms subsampling. Decoding options include CTC, TDT and attention encoder‑decoder, plus diarization via a Softformer bridge for target-speaker ASR. Training uses NeMo and pseudo-labeling, deployment runs on NVIDIA Triton + NIM microservices containerized for gRPC low‑latency and hundreds of concurrent streams, and a live demo shows robust noisy-song transcription.
AI-written summary · about
chapters (13)
- 0:00Nvidia speech AI models intro
- 0:42Model development focus areas
- 2:29Model architectures & decoding
- 4:33Fast conformer backbone
- 5:37Parakeet & Canary models
- 7:00Multi-speaker ASR
- 8:27Accuracy & readability models
- 9:53Hugging Face Open ASR leaderboard
- 10:55Noisy environment demo
- 11:17Training data & robustness
- 12:42Deployment with Reva & Nim
- 14:10Domain-specific customization
- 15:37Resources & getting started
related talks
25mBeyond Transcription: Building Voice AI That Understands Conversations
42mdata.bythebay.io: Shubho Sengupta,
36mNVIDIA NeMo: Toolkit for Conversational AI
14mLightning Talk: Enabling the Audio Modality for Language Models
23mNeMo: Open Source Toolkit for Conversational AI
40mKeynote: Open-source Multimodal AI in the Wild
43mTalk by Om Thakkar (OpenAI)
24mApplications of AI in news and media