Nvidia releases open-weight transcription AI that labels speakers in recorded audio

Nvidia released Nemotron 3.5 Transcribe on October 28, 2026, giving developers a free, open-weight model for turning recorded speech into text. Alongside it came the Granary v2 speech dataset and tools for adapting recognition to specialized vocabulary. The model combines transcription, translation, speaker labels and audio question answering, but does not support live streams.

One recording, several kinds of output

The model is available on Hugging Face and ModelScope. Nvidia says it handles recordings up to an hour long and labels up to eight speakers. Language coverage depends on the task:

• Transcription: speech-to-text support across 40 languages.
• Translation: 31 languages to or from English.
• Speaker attribution: labels identifying who spoke in 13 languages.

Developers steer the output with ordinary instructions. The same model can label speakers in a meeting, translate the recording into English, or answer a question about decisions made during the conversation. Audio question answering was trained in English, so the transcription language count should not be read as coverage for every function.

Prompts can also specify transcript style: preserve hesitations and spoken number forms, convert numbers into conventional notation, or remove disfluencies such as “uh.” For unfamiliar medical terms, Nvidia recommends supplying possible spellings as context, with instructions to use them only when the audio supports them.

Word-level timestamps are supported, too. Nvidia recommends checking names, speaker changes and answers against the recording rather than treating the generated text as an unquestionable record.

Accuracy claims and a carefully bounded speed test

Transcribe has 30 billion parameters, the model’s learned internal values, but activates about four billion per generated token. Nvidia says it runs on a single Nvidia GPU, from Jetson AGX Thor and DGX Spark systems to Blackwell data-center hardware. Developers can serve it through vLLM or run it locally with NeMo-Speech.cpp.

In its launch article, Nvidia reports first place among open-weight models on the English track of Hugging Face’s Open ASR Leaderboard, at 5.33% average word error rate. That measure counts transcription errors; lower is better. The cited English comparison excludes proprietary API models.

Speaker labeling uses cues passed into the language model alongside acoustic features, letting it assign turns during text generation. Nvidia reports an 18.26% speaker-labeling error rate on David AI’s 21-language benchmark, the lowest among 14 tested systems. That score measures mistakes in who spoke when, not just incorrect words.

Granary makes the training targets explicit

The companion Granary v2 release includes a reusable data-curation pipeline and a public manifest covering 463,062 hours across 48 languages and 16 source collections. Those figures describe the released dataset, not Transcribe’s complete training mixture. The pipeline prepared multilingual speech used to train the model.

Each recording receives three transcript forms corresponding to the different writing styles developers can request. The pipeline also produces audio captions, question-and-answer pairs grounded in speech, vocabulary hints and translations with automatic quality scores. Its checks screen for hallucination patterns and disagreement between transcripts before accepting records.

The public manifests include provenance and per-record quality flags. Source audio must be downloaded separately under each collection’s original license; developers must also follow the release’s published terms.

For specialized vocabulary that prompting does not resolve, Nvidia supplies NeMo Speech fine-tuning recipes. Its medical case study combines synthetic speech, accented English recordings and medical audio. Nvidia reports that fine-tuning cut word error rate on held-out speakers by 45.8% relative to the best prompt—a narrower finding than a general promise of medical accuracy.