LLMs

NVIDIA Nemotron 3.5 Transcribe unifies speech-to-text, diarization, translation, and audio QA in a single open-weight model

NVIDIA releases Nemotron 3.5 Transcribe, an open-weight model handling 40 languages, speaker attribution, and translation with top leaderboard performance and domain-adaptation tools.

Last verified:

NVIDIA Releases Nemotron 3.5 Transcribe with Unified Speech-Processing Capabilities

According to Hugging Face Blog, NVIDIA released Nemotron 3.5 Transcribe, an open-weight model consolidating transcription, speaker diarization, translation, and audio question-answering into a single checkpoint. The model handles recordings up to one hour long, supports transcription in 40 languages, performs speaker attribution in 13 languages, and translates 31 languages to or from English. Developers can control output using natural-language prompts and run the model on a single NVIDIA GPU, from edge systems like DGX Spark to Blackwell data-center platforms.

Leaderboard Performance and Accuracy Benchmarks

Nemotron 3.5 Transcribe ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for English-language tasks and occupies an accuracy-to-active-parameter frontier position on Artificial Analysis’s benchmark suite, according to the Hugging Face Blog. For speaker attribution, the model achieves the lowest diarization error rate across David AI’s 21-language leaderboard and ranks first in VoiceArena’s 14-system comparison. The model labels up to eight speakers simultaneously, enabling accurate attribution even in scenarios with overlapping dialogue.

Prompt-Based Output Customization and Domain Adaptation

A core differentiator of Nemotron 3.5 Transcribe is its ability to accept natural-language instructions that shape output without retraining. Hugging Face Blog details examples where the same checkpoint can handle distinct tasks—transcription with speaker labels, translation to English, or audio question-answering about meeting decisions—based on user prompts. For domain-specific vocabulary, developers can provide context in prompts (e.g., “Preserve standard medical spelling for terms like HbA1c and metformin”) before considering model fine-tuning.

Granary v2 and Fine-Tuning Recipes for Specialized Use Cases

The accompanying Granary v2 data pipeline and curation framework, released alongside Nemotron 3.5 Transcribe, integrates with NVIDIA’s NeMo Speech fine-tuning tools to enable model adaptation. According to Hugging Face Blog, a medical case study demonstrated that fine-tuning on domain-specific data using Granary v2 reduced word error rate by 45.8% relative to the base model on held-out speakers. This architecture allows organizations to combine prompt-level guidance with trainable adaptation, tailoring the model to their acoustic environment and terminology.

Why This Matters

Nemotron 3.5 Transcribe addresses a fragmentation problem in speech processing: teams building multilingual or multi-task speech applications have historically integrated separate models for transcription, diarization, and translation, introducing latency, memory overhead, and coordination complexity. Consolidating these capabilities in a single open-weight checkpoint reduces deployment friction and cloud inference cost, while prompt-based steering eliminates the need for retraining in many cases. The medical fine-tuning example (45.8% WER reduction) signals that organizations in regulated domains—healthcare, legal, financial services—can achieve production-grade accuracy on their own hardware and data without vendor lock-in. Teams evaluating speech infrastructure should benchmark Nemotron 3.5 Transcribe’s leaderboard parity against closed models like Gemini’s audio capabilities or Anthropic’s multimodal roadmap, particularly where domain adaptation and on-device or on-premise inference are requirements.

Frequently Asked Questions

What languages does Nemotron 3.5 Transcribe support?

The model transcribes 40 languages, translates 31 of them to or from English, and performs speaker attribution in 13 languages.

Can I adapt Nemotron 3.5 Transcribe to my domain or accent?

Yes. Developers can use natural language prompts to guide output during inference, or fine-tune the model on domain-specific data using Granary v2's curation pipeline and NeMo Speech tools. A medical case study showed 45.8% relative word error rate reduction on held-out speakers after fine-tuning.

What are the computational requirements?

According to Hugging Face Blog, the model runs on a single NVIDIA GPU, ranging from DGX Spark edge systems to Blackwell data-center infrastructure.

How does Nemotron 3.5 Transcribe handle overlapping speech?

The model includes speaker diarization capabilities that can label up to eight speakers in a single recording, handling overlapping dialogue in addition to sequential speech.

#speech-recognition #multilingual #open-weights #diarization #transcription