NVIDIA Nemotron 3.5 Transcribe is a single open-weight checkpoint that handles transcription, speaker attribution, translation, and audio question answering. It transcribes 40 languages, translates 31 of them to or from English, attributes speakers in 13, handles recordings up to an hour long, and labels up to eight speakers. Developers steer output with natural-language prompts — for example, asking for speaker labels, a translation, or decisions made in a meeting — and the same recording can yield different outputs depending on the instruction. Word-level timestamps are supported, while audio question answering is trained in English.

The model pairs a speech encoder with a hybrid Mamba–Transformer mixture-of-experts language model; only about 4B of its 30B parameters are active per token. It runs on a single NVIDIA GPU, from Jetson AGX Thor and DGX Spark to Blackwell data-center systems, and is built for recorded audio, not live streaming. On one RTX PRO 6000 Blackwell Server Edition GPU with NVFP4 precision, it processes 100 hours of audio in under three minutes on 30-second clips at high concurrency. It can be served with vLLM or run locally with NeMo-Speech.cpp.

Benchmark results reported in the article put it at the top of several open-weight rankings. On the Hugging Face Open ASR Leaderboard, it ranks first among open-weight models on English at 5.33% average WER, with 4.62% mean WER across seven languages and 10.98% on long-form. On Artificial Analysis’s Speech-to-Text Leaderboard, it ties for first among open-weight models at 2.8% AA-WER with about 4B active parameters, and scores 1.8% WER on the held-out AA-AgentTalk set versus 2.3% for the next-best open-weight model. For speaker attribution, it posts the lowest diarization error rate (18.26% DER across 21 languages) on David AI’s DAI-ASR-I18N diarization track and the lowest cpWER (32.4%) among 14 systems in VoiceArena’s Diarization-Bench, with fewer dropped words than the median system.

Granary v2 is the open data curation pipeline behind the model. Built with NVIDIA NeMo Curator, it generates three transcript forms per recording (verbatim text-normalized, inverse-text-normalized with disfluencies, and inverse-text-normalized without disfluencies), along with short audio captions, code-switching restoration, speech-grounded Q&A pairs, contextual ASR biasing terms, and translations with automatic quality scores. The public manifest covers 463,062 hours across 48 languages and 16 source corpora. The reusable pipeline has four stages — audio preparation, transcription with hallucination checks and fallbacks, task-label generation, and a HiFi filter that removes noisy, clipped, or multi-speaker segments; for VoxPopuli, HiFi reduced 354k hours to 111k hours across 22 European languages. The release describes the public dataset, not Transcribe’s full training mixture or sampling weights.

A medical ASR case study shows how prompt-level control extends to training. Starting with Eka Care’s open benchmark of Indian medical speech, a development-set prompt specifying English and the Latin alphabet cut full-benchmark WER from 30.8% to 23.8%. Fine-tuning with NeMo Speech — roughly 153 hours of synthetic speech, 29 hours of accented English from Common Voice, and 63 hours of medical recordings from MultiMed — brought full-set WER to 11.8%, about 62% fewer errors than the out-of-the-box model. Synthetic generation separates the written target from its spoken rendering: ‘HbA1c’ is the label while ‘H B A one C’ goes to the synthesizer. The final system uses one fixed convention prompt and one model pass, without retrieval or transcript post-processing.

The adapted model’s gains hold on held-out speakers. Fine-tuning reduces held-out WER from 22.44% for the best-prompt baseline to 12.17%, a 45.8% relative reduction; keyword WER falls from 9.2% to 6.5% and semantic WER from 10.0% to 7.6%. On the same held-out set, it beats an open medical-specialist model (13.79% WER) and a general-purpose ASR model (14.92%). General English retention is nearly unchanged: LibriSpeech test-clean WER rises from 1.54% to 1.69%. The medical specialist still wins on exact matches for brand-name drugs and diagnostics; the authors recommend real recordings of clinicians saying brand names as the next step, and at least 100 real labeled clips for checkpoint selection in any domain.

The recipe trains the speech encoder and projection and uses LoRA to adapt the language model; checkpoint selection is based on decoded development-set results rather than training loss alone. The full adaptation, from text generation through evaluation, took about 14 GPU-hours on one NVIDIA GB300. Developers can use the Granary v2 manifests or run the NeMo Curator pipeline with Ray and NVIDIA Dynamo backends on their own recordings or authorized public corpora. The release includes the model on Hugging Face and ModelScope, hosted endpoints from Baseten, DeepInfra, and Simplismart, the medical ASR case study with reproducible commands, fine-tuning tutorials for new-language and code-switching ASR, an agent skill for driving the workflow, and a ModelScope ms-swift path for multimodal fine-tuning. As with any speech model, the article advises spot-checking names, speaker changes, and answers against the recording.