NVIDIA open-sources Nemotron 3.5 Transcribe, tying for best open speech-to-text
On October 28, NVIDIA open-sourced the speech transcription model Nemotron 3.5 Transcribe (a 30B-total-parameter MoE with roughly 4B active) and released companion NeMo Speech open-weight ASR models. Both set open-source-leading results in third-party benchmarks and can already be deployed on vLLM 0.32.
Confirmed
- NVIDIA officially released Nemotron 3.5 Transcribe, an open-weight, promptable speech transcription model that combines multilingual transcription, speaker diarization, translation, and audio QA in a single model, supporting transcription in 40 languages and translation between English and 31 languages (per @NVIDIAAI).
- Artificial Analysis benchmarks show the model ties for first among open models for non-streaming transcription with a 2.8% WER, alongside a Mistral model.
- Artificial Analysis tested it on a DeepInfra-hosted endpoint: a 75x real-time speed factor, under 1 second to transcribe 1 minute of audio, at $2.00 per 1,000 minutes of audio.
- The vLLM team reports that NVIDIA NeMo Speech's open-weight English ASR model tops the Open ASR Leaderboard: 5.33% WER in English and an average 4.62% WER across 7 languages; the architecture is an audio encoder + Mamba–Transformer hybrid MoE, deployable on vLLM 0.32.
Why it matters
- A single model covering transcription, speaker diarization, translation, and audio QA, combined with open weights and extremely low cost ($2 per 1,000 minutes), dramatically lowers the barrier to high-quality speech transcription.
- The highly efficient MoE design (4B active out of 30B total parameters) delivers 75x real-time speed, demonstrating MoE's engineering viability for the speech modality, and native vLLM support means the community can self-deploy quickly.
2026-10-28 ~ 2026-10-28 · 9 related posts
- Episode 1: NVIDIA open-sources Nemotron 3.5 Transcribe, tying for best open speech-to-text(2026-10-28, 9 posts)
- Episode 2: vLLM Reviews Native Support for NVIDIA Nemotron-3.5-Transcribe(2026-10-28, 2 posts)
Primary sources
- NVIDIA open-sources Nemotron 3.5 Transcribe: 30B MoE ASR model with ~4B active params — NVIDIAAI ·
- NVIDIA open-sources Nemotron 3.5 Transcribe 30B ASR model; fine-tuning cuts word error rate 45.8% — NVIDIAAI ·
- Nemotron 3.5 Transcribe hits 75x speed factor at $2 per 1,000 minutes in AA benchmark — ArtificialAnlys ·
- NVIDIA's Nemotron 3.5 Transcribe ties for #1 open-weight STT at 2.8% WER with just 4B active params — ArtificialAnlys · 2026-10-28
- [source] Nemotron 3.5 Transcribe hits 75x speed factor at $2 per 1,000 minutes in AA benchmark — ArtificialAnlys · 2026-10-28
- [source] NVIDIA open-sources Nemotron 3.5 Transcribe: 30B MoE ASR model with ~4B active params — NVIDIAAI · 2026-10-28
- [source] NVIDIA open-sources Nemotron 3.5 Transcribe 30B ASR model; fine-tuning cuts word error rate 45.8% — NVIDIAAI · 2026-10-28
- NVIDIA releases open Nemotron 3.5 Transcribe 30B A4B speech transcription model on Hugging Face — coder543 · 2026-10-28
- NVIDIA NeMo Speech hits 5.33% English / 4.62% multilingual WER among open models — vllm_project · 2026-10-28
- NVIDIA open-sources Nemotron 3.5 Transcribe: 40-language speech model topping Open ASR — realmrfakename · 2026-10-28
2 near-duplicate retellings: vllm_project · baseten