Nvidia released Nemotron 3.5 Transcribe on Wednesday (October 28), an open speech recognition model with 30 billion parameters (4.3 billion active thanks to a hybrid Mamba-Transformer MoE architecture) that transcribes audio in up to 40 languages, translates 31 languages to and from English, and identifies up to eight different speakers in a single recording. The model is available under the OpenMDW-1.1 license, with weights on Hugging Face and fine-tuning code in the NeMo Speech repository.
What sets Nemotron 3.5 Transcribe apart is not just the number of languages — smaller models already cover dozens of tongues — but the combination of capabilities that previously required three or four separate models: speaker-attributed transcription (diarization), translation, language identification, contextual transcription, and audio question answering, all in a single checkpoint. Nvidia calls this "intelligent transcription," and the concept makes sense when you look at the benchmarks.
On Hugging Face's Open ASR Leaderboard, the model achieved 4.71% Word Error Rate (WER) across the nine public English test sets and 4.62% on the multilingual leaderboard. For context, the next best open-weight model sits at 4.94% (Qwen3-ASR-1.7B), but while Qwen operates with 1.7 billion parameters against Nvidia's 30 billion, Transcribe's active parameter count of 4.3 billion brings it closer to Qwen3's tier. On the multilingual subset, the advantage is even clearer: 4.62% versus 6.49% for Qwen.
But the number worth paying attention to is the long-form audio score. On the long-form leaderboard sets, Transcribe hit 10.98% WER against 13.37% for Parakeet-TDT-0.6B-v3. Transcribing an entire hour of meeting audio with multiple speakers while maintaining that level of accuracy requires more than a model that's good at short clips — it demands context memory and the ability to model how speakers change their tone over time. Transcribe succeeds here because it directly injects speaker activity (via the frozen Nemotron 3 Diarization model) into the encoder representations, allowing the LLM to produce speaker-labeled transcripts without needing a separate cascade system.
Nvidia also highlighted a six-fold reduction in named entity errors on the ContextASR-Bench. This means that when given a list of expected names or terms in the prompt, the model dramatically reduces errors like "Stefan" becoming "Steven" or "NVIDIA" becoming "new di." In practice, this is crucial for medical transcription (drug names, procedures) and legal transcription (parties' names, technical terms) — two markets where entity accuracy matters as much as the transcript itself.
In terms of efficiency, the model is served via Nvidia's NIM and supports native speculative decoding via Multi-Token Prediction (MTP). With the NVFP4 checkpoint and DFlash2 draft model, throughput reaches 2,238 seconds of audio processed per second on a single RTX PRO 6000 Blackwell Server Edition GPU. Nvidia also provides weights in BF16, FP8, and NVFP4, plus GGUF files for NeMo-Speech.cpp, enabling local execution on edge devices like the Jetson AGX Thor, the DGX Spark, and — soon — quantized versions for consumer RTX GPUs.
Nvidia is not alone in this race. October has been particularly busy in the voice AI segment: Mistral shipped Voxtral 3 earlier in the month, ElevenLabs launched Realtime for instant voice agents, and Salesforce presented Launa-Omni. But Transcribe's difference is that it doesn't try to do everything — it focuses on recorded audio, not live streams. That limitation is intentional: the model processes entire recordings, which allows using full context for better accuracy, but it excludes real-time transcription for calls and live captions, where other models hold an advantage.
The question is whether the voice AI industry is fragmenting too much — specialized models for transcription, diarization, translation, and audio QA versus omni models that attempt to do it all. Transcribe shows that for well-defined use cases (post-call analytics, medical transcription, meeting tools), the specialized approach still wins on accuracy. The question is not which model is most versatile, but which is most useful for the specific problem you're solving.
Sources: NVIDIA Research, TechCrunch, Hugging Face Blog
✓ Independent sources cross-checked and verified before publishing
