Nvidia Releases Nemotron 3.5 Transcribe, a 34B Offline Speech Model

Nvidia released Nemotron 3.5 Transcribe on Oct. 28, an open-weight speech model with 34 billion parameters, of which 4.3 billion are active, covering transcription in 40 languages. A benchmark run by the voice-agent testing firm Cekura put it first of 21 models on word error rate, though Cekura called its real-call sample small.

The model card, published on Hugging Face, calls it "an open-weight, offline speech and audio model." It says the weights are governed by the OpenMDW License Agreement, version 1.1, and that the model "is ready for commercial or non-commercial use." The card does not say whether access is gated, and The Terminal did not read the licence text itself.

How the model is built

The card describes a Mamba-2 and Transformer hybrid mixture-of-experts language model, which routes each token through a subset of its parameters, paired with an audio encoder and a speaker encoder. The 34 billion figure includes the audio encoder. Training used about 1.72 million hours of unique audio, roughly 372 million examples, plus 5.6 billion text tokens, drawn from sources including Granary v2, AMI, AISHELL-4 and LibriSpeech.

The reference deployment is one H100 with 80 GB of memory. The model collection page describes support for one hour of audio and a 64,000-token context.

Another open-weight mixture-of-experts release on the site's tracker is Mistral Large 4, which needs an eight-GPU node.

What Nvidia and Cekura measured

The card's figures are Nvidia's own. On the English Open ASR mean, word error rate, where lower is better, is 4.71 against 4.94 for Qwen3-ASR-1.7B, the next best open-weight model it lists. On long-form audio the card gives 10.98 against 13.37 for Parakeet-TDT-0.6B-v3, and on a seven-language mean 4.62 against 6.49.

Test, vendor-supplied	Nemotron 3.5 Transcribe	Next best listed
English Open ASR mean	4.71	4.94 (Qwen3-ASR-1.7B)
Long-form Open ASR mean	10.98	13.37 (Parakeet-TDT-0.6B-v3)
Multilingual mean, 7 languages	4.62	6.49 (Qwen3-ASR-1.7B)
Multi-speaker cpWER	15.80	16.83 (DiCoW v3.3 + DiariZen)

Cekura's post, dated Oct. 29, is the outside check. On its Converse-STT benchmark, which uses 1,000 clips and eight real phone-call recordings across 21 models, Nemotron 3.5 Transcribe scored 2.28 percent overall, the lowest. Zoom Scribe Pro v2.1 was next at 2.35 percent.

1. Nemotron 3.5 Transcribe2.28 %
2. Zoom Scribe Pro v2.12.35 %

Source: Cekura, Nemotron 3.5 Transcribe post dated 2026-10-29, accessed 2026-11-03

On the eight real calls alone it scored 3.03 percent, and Reson8 was lower at 2.93 percent. Cekura said eight recordings is a small sample and called the real-call ranking a close race.

Limits both sources state

The card lists its own error types: substitutions, deletions, repetitions, hallucinated speech, wrong language choice and speaker attribution errors. It says text-only performance is weaker than audio performance, and that spoken instruction-following and audio reasoning were evaluated only in English. It states no maximum speaker count and no real-time factor.

Cekura said "This is an offline model" and that "Live voice agents still need a streaming model." It cites an RTFx of 462 from the Open ASR Leaderboard, which is not its own measurement, and it ran no hallucination test.

The Terminal's earlier report on Nvidia's auto-ontology release covered a different Nvidia project. Cekura recommends testing on a team's own calls before switching models, and the card recommends targeted tests on long recordings, overlapping speakers, noise and code-switching.