LLMs

Nvidia releases Nemotron 3.5 Transcribe, a 30B-parameter open-weight speech model supporting 40 languages

Nvidia's free transcription model handles multilingual audio up to 1 hour with speaker diarization, targeting business and medical workflows.

Last verified:

Nvidia released Nemotron 3.5 Transcribe, a free 30-billion-parameter open-weight speech model, on October 28, 2026. According to TechCrunch AI, the system transcribes audio in up to 40 languages, performs speaker diarization for eight speakers, and handles recordings up to one hour in length. The model also translates between 31 languages and English, positioning it as a multilingual alternative to closed commercial transcription services.

Technical Capabilities and Benchmarking

The model achieves competitive performance on industry benchmarks. TechCrunch AI reports that Nemotron 3.5 Transcribe ranks favorably against rivals on Hugging Face’s Open ASR Leaderboard, though the publication does not specify which benchmarks or by what margin. The 30B-parameter architecture allows deployment on a single accelerator—Nvidia’s Jetson AGX Thor robotics platform, DGX Spark chip used in consumer laptops, and enterprise Blackwell datacenter GPUs all support the model natively. This single-GPU constraint differs from larger enterprise models that require multi-node clusters, though it remains a practical limitation for production inference on entry-level hardware.

Nvidia also released version 2 of Granary, its open-source speech dataset, alongside the model. The company positions both assets as customization tooling for developers integrating industry-specific terminology into transcription pipelines—a workflow TechCrunch AI identifies as especially relevant for business meeting transcription and medical dictation use cases.

Competitive Landscape and Market Timing

The Nemotron 3.5 release lands in an accelerating voice-AI release cycle. TechCrunch AI notes that Mistral shipped Voxtral 3 earlier in October 2026, ElevenLabs launched Realtime for low-latency voice agents, and Salesforce released Launa-Omni, its proprietary voice model. The proliferation signals that voice has become a primary interface vector in enterprise AI—moving beyond chatbot integrations into live interaction and specialized workflows.

One functional gap: TechCrunch AI points out that Nemotron 3.5 Transcribe handles only pre-recorded audio, not live streams. This excludes real-time meeting transcription and call-center automation use cases where latency matters, though batch transcription for meeting archives, medical notes, and legal depositions remains a large addressable market.

Why This Matters

Open-weight transcription models lower the barrier to deploying ASR in regulated industries where data residency or model auditability constraints prohibit cloud APIs. Teams building medical transcription or legal discovery tools—sectors with strict data governance—can now run Nemotron 3.5 on-premise without vendor lock-in. For developers currently licensing commercial ASR (Amazon Transcribe, Google Cloud Speech-to-Text, AssemblyAI), this release creates an option to reduce per-minute API costs, particularly at scale. The single-GPU deployment profile also makes the model accessible to resource-constrained startups or edge deployments in remote clinics or field offices where cluster infrastructure is unavailable. However, the lack of streaming support limits applicability to live-interaction use cases, meaning commercial APIs remain entrenched in real-time applications.

Frequently Asked Questions

What languages does Nemotron 3.5 Transcribe support?

The model transcribes audio in up to 40 languages and translates between 31 languages and English.

Can Nemotron 3.5 Transcribe run on consumer hardware?

Yes. According to TechCrunch AI, the 30B-parameter model runs on single-chip systems including Nvidia's Jetson AGX Thor and DGX Spark laptops, as well as datacenter GPUs like Blackwell.

Does this model support live audio streams?

No. TechCrunch AI notes that Nemotron 3.5 Transcribe works only with recorded audio, not real-time streaming.

How many speakers can the model distinguish?

The model supports speaker diarization for up to eight distinct speakers in a single recording.

#voice-ai #open-weights #transcription #multilingual #nvidia