Nvidia Releases Free Nemotron 3.5 Transcribe AI Model Supporting 40 Languages

Nvidia Launches Comprehensive Open Transcription Model

Nvidia released Nemotron 3.5 Transcribe on Wednesday, a free open-weight AI model designed to transcribe audio in 40 languages and translate 31 languages to and from English1

1

2

2

. This open speech model represents a significant advancement in multilingual audio transcription, consolidating multiple speech processing tasks into a single model pass. Unlike traditional transcription workflows that separately detect speech, transcribe it, identify speakers, and answer questions, Nemotron 3.5 Transcribe handles all these functions simultaneously3

3

.

Technical Architecture and Performance Benchmarks

The model features a 30-billion parameter mixture-of-experts model architecture with only 4 billion active parameters per step, keeping it fast and cost-effective to run2

2

3

3

. Nvidia claims the open transcription model ranks first among open models on Hugging Face's English speech-recognition leaderboard, achieving an average word error rate of 5.33%2

2

. Out of the box, the model hits 30.8% error rates, which can be reduced to 20.8% with careful prompting3

3

. The model handles recordings up to an hour long and performs speaker diarization for up to eight speakers1

1

2

2

.

Hardware Flexibility and Deployment Options

Nemotron 3.5 Transcribe runs on a single Nvidia chip across a wide range of hardware configurations, from edge devices like Jetson AGX Thor robotics controllers and DGX Spark desktop AI supercomputers to Blackwell data-center GPUs1

1

3

3

. A Blackwell Server Edition GPU with NVFP4 precision and one-quarter model compression can process 100 hours of audio in under three minutes3

3

. This flexibility allows developers to deploy the model anywhere from the edge to the data center depending on compression needs and performance requirements.

Fine-Tuning Capabilities for Domain-Specific Applications

Developers control the open-weight AI model with plain-language prompts, requesting clean or word-for-word transcripts, or providing medical or product terms to listen for2

2

. Users can instruct the model with prompts like "Transcribe the recording and label each speaker," "Translate this audio into English," or "What decisions were made in this meeting?"3

3

. The model's fine-tuning capabilities shine in specialized applications. When trained on medical transcription data from Indian doctors, the model cut errors by 45.8% on held-out test clips, achieving an 11.8% word error rate on the Eka Care Indian medical speech benchmark2

2

3

3

. This fine-tuning took approximately 14 hours on one GPU2

2

.

Granary v2 Dataset Expansion

Alongside Nemotron 3.5 Transcribe, Nvidia released Granary v2, a free dataset of 463,062 hours of speech across 48 languages, up from 25 languages in the first version2

2

. The dataset actually covers 843,579 hours across 48 languages and 19 source corpora, up from about 643,000 hours3

3

. Built with Nvidia NeMo Curator, Granary v2 generates three transcripts for each recording in clean verbatim styles: one providing everything including "ums" and "ahs," another reducing full wording to figures like spelled-out numbers to symbols, and a third eliminating filler words3

3

. Developers download the audio from each original source under its own license2

2

.

Market Positioning and Competitive Landscape

The release targets developers building applications for business meetings and medical transcription, two of AI's most established footholds in the workplace1

1

. While competitors like SpaceXAI's Grok transcription charge 10 cents per hour, Nemotron 3.5 Transcribe is free for commercial use2

2

. The release comes during a busy period of AI voice releases, with Mistral shipping Voxtral 3 earlier this month, ElevenLabs launching Realtime for instant speech in voice agents, and Salesforce releasing Launa-Omni1

1

. The model joins Nvidia's open-weight voice models, including the compact Parakeet at 1.1 billion parameters1

1

. One current limitation is that Transcribe works with recorded audio, not live streams1

1

2

2

. Transcribe is available as a Hugging Face demo, where developers can download a checkpoint and adapt it for domain expertise3

3

.