On October 28, 2026, NVIDIA released the speech model Nemotron 3.5 Transcribe (30B-A4B), which performs speech transcription, translation, speaker diarization, and audio question-answering in a single checkpoint across 40 languages, including Russian. In the open ASR leaderboard on Hugging Face, the model took first place among open models, and the dataset Granary v2, containing 463,000 hours of labeled speech in 48 languages, was published alongside it.

image
image

What happened

The model is built on a hybrid scheme: a speech encoder plus a language model consisting of Mamba-2 and Transformer layers with mixture-of-experts; it has a total of 34 billion parameters, of which approximately 4.3 billion are active per token. The checkpoint supports 40 languages, including Russian, translates between 31 languages and English in both directions, distinguishes up to 8 speakers, and accepts recordings up to one hour long. In the open ASR leaderboard on Hugging Face, Nemotron 3.5 Transcribe took first place among open models: 5.33% WER on English, 4.62% on the multilingual part, and 10.98% on long recordings; the Artificial Analysis ranking cites a value of 2.8% WER with approximately 4 billion active parameters. In parallel, the dataset Granary v2 was released, containing 463,062 hours of speech in 48 languages, collected from 16 open corpora: the repository publishes manifests and the processing pipeline, while audio is downloaded separately according to the licenses of the source corpora.

Context

Prior to this release, speech stacks were typically assembled as a cascade: separate transcription, separate translation, and separate diarization, meaning teams either paid for multiple services or manually stitched together several models. Nemotron 3.5 Transcribe belongs to the class of instruction-following speech models: tasks are selected via instruction rather than by switching tools. The choice of a Mamba-2 and Transformer hybrid with mixture-of-experts is specifically designed for long recordings, where the quadratic cost of transformer attention would become a bottleneck, while the expert mechanism allows maintaining a large total parameter volume with a small active budget per token. The second part of the release, the Granary v2 dataset, is not just an audio archive but labeled data with a reproducible open pipeline, making it a set suitable for fine-tuning third-party models for specific domains.

Why this matters for the industry

The main claim to the industry is efficiency: according to Artificial Analysis, the model achieves 2.8% WER with approximately 4 billion active parameters, whereas the nearest competitors in this ranking consume about 24 billion active parameters. A single checkpoint instead of a cascade of two to three models simplifies and reduces the cost of speech products—from meeting and call bots to media archive processing; the optimized NIM build processes a 30-second recording in 0.13 seconds, making batch processing cheap on server deployments. Granary v2 with an open pipeline provides companies with a ready-made base for fine-tuning ASR and translation models for their own vocabulary. If the leaderboard figures are confirmed by independent checks, a wave of such domain fine-tunings and price pressure on commercial ASR APIs from local deployments is likely.

Why this matters for users

The model can be tried right now: weights are published on Hugging Face and ModelScope under the OpenMDW-1.1 license, there is a demo on Hugging Face Spaces, and local deployment is described via vLLM and NeMo Speech. The entry barrier is a minimum of one H100 80GB accelerator, meaning this is a server scenario, not a laptop one. For Russian-language recordings, all stated tasks are available: transcription, translation, speaker diarization, and switching between languages within a single phrase (code-switching); batch processing of files immediately outputs text with timestamps and speaker labels, which is suitable for transcribing interviews, meetings, and call recordings.

What is currently unknown / limitations

There is a nuance with the quality figures: the open Hugging Face leaderboard lists 5.33% WER on English, while Artificial Analysis cites 2.8%—these are different rankings with different test sets, so correct comparison is currently difficult, and there are no independent reproductions yet. Quality noticeably drops on long recordings: 10.98% WER on the long part of the ranking shows that support for one-hour input does not mean uniform quality across the entire length. Audio question-answering works only in English, and streaming recognition is not supported—the model accepts only ready-made files. In Granary v2, manifests and the pipeline are distributed, while the audio itself must be downloaded separately, taking into account the licenses of the source corpora.

Sources

Author

Look at AI, editorial team