NVIDIA NeMo Speech tops open ASR at 5.33% WER, servable via vLLM 0.32 plugin
vllm_project · x · 2026-10-28
NVIDIA's NeMo Speech claims #1 among open-weight models on English Open ASR with 5.33% WER, and 4.62% mean WER across 7 languages for multilingual Open ASR. The architecture pairs an audio encoder with a hybrid Mamba–Transformer MoE, keeping 4.3B parameters active including the encoder.
Deployment: the NeMo Speech vLLM 0.32.0 plugin serves checkpoints with OpenAI-compatible chat completions and MTP speculative decoding, with a full guide. Docs also cover CTC word-timestamp alignment (single-pass banded aligner, configurable Sortformer speaker log-prob prior, diarization-assisted boundary refinement capped at 12 frames) and speaker mapping details.
Related event: NVIDIA Open-Sources Nemotron 3.5 Transcribe 30B Speech Model(10 posts)→
More from Infra
- KV cache billing dispute: realized-hit pricing may incentivize gaming, OpenRouter cited — charles_irl · 2026-10-29
- PyTorch 2.15 ships with default NCCL rewrite, NVFP4 decode on Blackwell — PyTorch · 2026-10-29
- US data center construction hits $85B annual pace, up 73% year over year — BenBajarin · 2026-10-29
- Azure Cosmos DB recipe: isolated disposable databases for a swarm of parallel coding agents — adnan_hashmi · 2026-10-29
- Short seller claims Cerebras wafers degrade after ~2 years, undermining its business — himanshustwts · 2026-10-29
- Apple's RAM upgrade pricing: 512GB costs a staggering $13,950 over base — firstadopter · 2026-10-29