HeyHomees Lab Notes · No. 14 · NVIDIA · Published November 2, 2026
By HeyHomees Newsroom ·
In 30 seconds
- 34B total parameters (4.3B active) enable high-accuracy transcription across 40 languages.
- Mean WER 4.71 on English ASR and 4.62 on multilingual ASR on the Open ASR Leaderboard.
What is it?
Nemotron-3.5-Transcribe is an offline speech and audio model that converts recordings and optional text prompts into transcribed text. It supports transcription, translation, speaker attribution, timestamping, language identification and audio question answering. The model is released under the OpenMDW-1.1 license and is ready for commercial or non-commercial use.
How does it work?
The model uses a Mamba-2–Transformer hybrid Mixture-of-Experts architecture with a Parallel Expert Encoder that includes a Transformer ASR encoder and a speaker encoder. Approximately 4.3B parameters are active during inference, and the model can be run in BF16, FP8 or NVFP4 precision. It supports speculative decoding via native Multi-Token Prediction and optional DFlash2 draft model.
How well does it do?
On the Open ASR Leaderboard, Nemotron-3.5-Transcribe achieves a mean English WER of 4.71, a multilingual mean WER of 4.62 across seven languages, and a long-form mean WER of 10.98. For multi-speaker ASR it records a mean cpWER of 15.80, and style-controlled WERs of 7.56 (default) and 6.16 (instructed).
What does it need to run?
The reference deployment runs on a single NVIDIA H100 80 GB GPU. An RTX PRO 6000 Blackwell GPU can achieve up to 2,238 seconds of audio per second using NVFP4 precision with DFlash2 speculative decoding, though this speed is from a non-open-source vLLM integration.
By the numbers
- Total parameters: 34B (4.3B active)
- Context length: 65,536 tokens; recordings up to 1 hour
- Supported languages: 40
- RTX PRO 6000 throughput: 2,238 seconds of audio per second
The honest read
The reported error rates (e.g., 4.71 % English WER) are self-reported by NVIDIA on specific public benchmark sets and do not cover all possible use cases. Efficiency numbers such as 2,238 seconds per second are from an optimized vLLM integration that is not yet part of the open-source release, so real-world performance may be slower. The model’s 34B parameter size includes many inactive experts, and actual active compute is 4.3B parameters, which limits direct comparison with dense models. The benchmarks focus on mean WER and cpWER; they do not measure latency for streaming or real-time scenarios, and the model is described as offline, meaning it is not a streaming ASR service.
What we don't know yet
- Exact download size of the model files.
- Performance on languages beyond the seven evaluated in the multilingual leaderboard.
- Real-time streaming capabilities or latency on standard hardware.
Sources
- Model card, Hugging Face (primary)
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.