Summary
NVIDIA has released Nemotron 3.5 Transcribe, an open-weight model for offline speech transcription and related audio tasks in 40 languages. It supports recordings up to one hour and can produce speaker labels and word timestamps.
NVIDIA released Nemotron 3.5 Transcribe on October 28, 2026, describing it as an open-weight model for offline speech and audio processing. It can transcribe recordings in 40 languages and also supports speech translation, speaker-attributed transcripts, word timestamps, language identification and audio question answering. NVIDIA says the model can process recordings up to one hour long, with a context length of 65,536 tokens.
What the model can do
The model takes audio along with text instructions and optional context, such as vocabulary hints. That lets developers request different transcript styles: verbatim wording with fillers and repetitions, or cleaner text that omits them. For recordings with multiple speakers, it can assign speaker labels and, when configured, return word-level start and end times.
Those labels distinguish speakers within a recording; NVIDIA says they do not identify people in the real world. The model card also notes that punctuation, normalization and speaker attribution depend on the task and prompt. Although its audio encoder processes recordings in internal 30-second chunks, NVIDIA describes the model as an offline system, not a real-time streaming transcription service.
The language matrix lists English, including Indian English, and Hindi, Bengali, Marathi, Tamil and Telugu among its supported transcription languages. Support varies by task and language: the card lists translation between Hindi and English, but does not list that capability for Bengali, Marathi, Tamil or Telugu.
Architecture, benchmarks and deployment
Nemotron 3.5 Transcribe has 34 billion total parameters, with about 4.3 billion active, including its audio encoder. It combines a Mamba-2–Transformer mixture-of-experts language model with audio and speaker encoders. In a mixture-of-experts model, only selected parts of the larger network are active for a given input. NVIDIA provides BF16 weights as well as FP8 and NVFP4 variants. Its deployment guide describes running the model on one H100 GPU with 80 GB of memory; supported hardware and precision vary across other configurations.
NVIDIA’s model card reports a 4.71% mean word error rate across nine English ASR test sets and a 10.98% mean across three long-form benchmarks. For the Multilingual Open ASR Leaderboard, it reports 4.62% across seven languages; Hindi in that evaluation uses orthographically informed WER. On the Monsoon English (India) benchmark, the card lists 3.65% WER. These are results reported by NVIDIA, and scores should be read within their benchmark conditions rather than treated as a single measure of performance across languages and tasks.
The model is released under the OpenMDW-1.1 license. NVIDIA says it is intended for commercial and non-commercial use, subject to that license and any applicable third-party terms. The model card describes audio understanding, reasoning and spoken-instruction evaluations as English-only, so those capabilities may not work reliably in other languages.