Nvidia Opens Nemotron 3.5 Transcribe for Multilingual Audio, Speaker Labels and Custom Speech Workflows

Nvidia has released Nemotron 3.5 Transcribe, an open-weight speech model that combines transcription, translation, speaker attribution and audio question answering in one system. The model is available through Hugging Face and ModelScope, alongside the Granary v2 speech dataset and tools for adapting speech applications.

Nemotron 3.5 Transcribe processes recordings lasting up to 1 hour, supports 40 languages for transcription and translates 31 languages to or from English. It can identify as many as 8 speakers, giving developers a single model for meeting records, interviews, medical dictation and other audio workflows. TechCrunch describes the release as free and open, while Nvidia’s product announcement emphasizes that developers can guide output with natural-language prompts.

The model’s main distinction is its combination of speech recognition and post-transcription tasks. A developer can ask for a transcript, request speaker labels, translate speech, or query the contents of an audio file without stitching together separate systems for each job.

Nemotron 3.5 Transcribe also supports formatting instructions and domain-specific vocabulary. That matters in fields such as medicine, where a transcript can be technically accurate at the word level yet still mishandle drug names, abbreviations or specialized terminology. Nvidia says its medical fine-tuning example reduced word error rate by 45.8% relative to the strongest prompt-only configuration on held-out speakers. The Hugging Face announcement also places the model first among open-weight systems on the English Hugging Face Open ASR Leaderboard and reports leading diarization results on other evaluations.

Nemotron 3.5 Transcribe is designed for recorded audio, not live streams. That limits its direct use in real-time calls, live captions and voice agents that must respond while a person is speaking, although recorded meetings and uploaded clinical audio fit the model’s intended workflow.

The hardware footprint gives the release a wider range of deployment options. Nvidia says the 30-billion-parameter model can run on a single GPU, from DGX Spark systems to Blackwell data-center hardware. TechCrunch also identifies Nvidia’s Jetson AGX Thor robotics platform as a target device and places the model alongside the company’s smaller Parakeet 1.1B multilingual speech model. The combination of open weights and single-chip execution gives developers more control over where audio is processed, including private infrastructure.

Granary v2 gives developers data and a reusable curation pipeline for adapting speech systems to specific accents, languages and terminology. Nvidia has published the dataset on Hugging Face and ModelScope, while its NeMo Speech tools provide recipes for fine-tuning Nemotron 3.5 Transcribe.

That training path extends the model’s prompt-based control. Prompts can tell the system how to format a transcript or spell a term, but fine-tuning can teach it recurring patterns in a company’s audio. The medical example provides a concrete measure of the difference: Nvidia reports a 45.8% relative word-error-rate reduction after adaptation. For businesses, the attraction is a speech stack that can be tuned without sending sensitive recordings to an external API, provided the organization has the required hardware and engineering capacity.

Nvidia’s release enters a crowded voice market. Mistral shipped Voxtral 3 earlier in October, ElevenLabs introduced Realtime for voice agents, and Salesforce released Launa-Omni, according to TechCrunch. Those launches divide attention between transcription accuracy, real-time interaction and broader voice-agent capabilities.

Nemotron 3.5 Transcribe’s strongest position is its open, multilingual workflow for recorded speech. Its benchmark claims matter because transcription quality, diarization and translation often break down differently across languages and domains. The model’s limits also define its market role: it addresses post-processing and analysis of audio, while real-time products compete on latency and turn-taking. Developers choosing among these systems will weigh openness, deployment control, language coverage, customization and whether their application needs live audio.

Developers can now test Nemotron 3.5 Transcribe, Granary v2 and NeMo Speech without building a speech pipeline from separate recognition, diarization and translation components. That lowers integration work for meeting intelligence, medical transcription, multilingual archives and audio search, though operating a 30-billion-parameter model still requires suitable GPU capacity.

The next practical test is adoption outside Nvidia’s benchmarks. Accuracy across accents, noisy rooms, medical terminology and less-represented languages will determine whether the model becomes a dependable production tool. Its open weights and training resources give teams a path to local deployment and customization, while the absence of live-audio support leaves real-time voice applications to newer specialized systems.