Description
NVIDIA Nemotron 3.5 Transcribe is a sophisticated automatic speech recognition (ASR) model, featuring 34 billion parameters (4.3 billion active) and incorporating an audio encoder. Its architecture is a Mamba-2–Transformer hybrid Mixture of Experts language model, enhanced with a Parallel Expert Encoder for ASR and speaker recognition. The model is designed for high accuracy, customization, and fine-tuning, making it a powerful tool for developers and researchers.
Read More
This model supports a wide array of tasks, including verbatim and written form transcription, contextual transcription, speaker-attributed transcription, rich transcription, speech translation, language identification, and audio question answering. It is proficient in 40 languages, with varying quality levels from 'Leading' to 'Supported' across different locales and capabilities like ASR, translation, multispeaker ASR, and code-switching.
Nemotron 3.5 Transcribe is optimized for deployment on NVIDIA hardware, specifically requiring 1× H100 80 GB for optimal performance. It supports BF16 precision, with FP8 and NVFP4 variants also available. The model can process recordings up to 1 hour long and has a context length of 65,536 tokens. Inputs include audio (WAV, FLAC, MP3, OGG) and text, generating text outputs with optional word timestamps.
The model demonstrates leading performance across various benchmarks, including multi-speaker ASR, style control, English ASR, long-form ASR, and multilingual ASR. Its prompt guide offers flexible options for different use cases, such as generating research transcripts, readable text for captions, transcribing domain-specific vocabulary, identifying speakers, translating speech, and answering questions from audio recordings.
NVIDIA Nemotron 3.5 Transcribe is licensed under OpenMDW-1.1, ensuring clear terms of use. It is intended for developers and researchers building speech and audio applications that require application-specific evaluation and review. The model's robust capabilities and extensive language support make it a versatile solution for complex audio processing and understanding tasks.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsLinux, Windows
NVIDIA Nemotron 3.5 Transcribe Highlights
34 Billion Parameters (4.3B active) with audio encoder
Mamba-2–Transformer hybrid Mixture of Experts architecture
Supports BF16, FP8, and NVFP4 precision
Context length of 65,536 tokens for recordings up to 1 hour
Processes audio and text inputs, outputs text with optional word timestamps
Supports 40 languages for transcription and translation
Performs verbatim and written form transcription
Offers contextual and speaker-attributed transcription
Includes rich transcription and speech translation capabilities
Features language identification and audio question answering
Optimized for NVIDIA H100 80 GB hardware deployment
Native Multi-Token Prediction (MTP) for speculative decoding
Configurable reasoning mode via chat template
Achieves high accuracy across multiple ASR benchmarks
OpenMDW-1.1 License
Getting Started with NVIDIA Nemotron 3.5 Transcribe
Access Model: Obtain access to the NVIDIA Nemotron 3.5 Transcribe model via Hugging Face.
Set Up Environment: Configure your environment with necessary NVIDIA hardware (e.g., H100 80 GB) and inference runtimes (vLLM, Hugging Face Transformers).
Prepare Inputs: Provide audio files in WAV, FLAC, MP3, or OGG formats, and/or text prompts.
Formulate Prompts: Use recommended prompt structures for specific tasks like verbatim ASR, clean transcripts, speaker attribution, or speech translation.
Send Request: Send requests to the model server, specifying decoding modes and reasoning settings.
Process Output: Receive generated text outputs, which may include transcripts, translations, speaker tags, or answers to audio questions.
Integrate Software: Integrate the model into applications using supported architectures and operating systems (Linux, Windows via WSL2).
NVIDIA Nemotron 3.5 Transcribe's Use Cases
- Research Transcripts
- Captions and Notes
- Technical Discussions Transcription
- Meeting Review Workflows
- Dictation and Voice Interfaces
- Multilingual Media Localization
- Audio Search and Understanding
- Code-Switching Transcription
- Structured Value Recovery
- Spoken Assistant Requests







