23 Daily — According to its Product Hunt page, NVIDIA introduced Nemotron 3.5 Transcribe on October 28, 2026. This open speech large language model (LLM) enables real-time transcription in 40 languages, supporting both verbatim and written formats. It features context understanding, speaker attribution, and word-level timestamps. Nemotron 3.5 is ranked as the best open-weights model on the English and multilingual Open ASR Leaderboards and allows transcription style selection via natural language instructions without retraining.
Technical Architecture and Key Features of Nemotron 3.5
Nemotron 3.5 Transcribe integrates a Transformer speech encoder and a frozen Nemotron 3 Diarization model with the Nemotron 3.5 Lightning 30B-A3B hybrid Mamba–Transformer mixture-of-experts LLM, totaling approximately 4.3 billion active parameters. This single model performs verbatim, written-form, contextual, speaker-attributed, and rich transcription, supports speech translation and language identification across 40 languages, and processes recordings up to one hour long. It also optionally outputs speaker-attributed word-level timestamps. Transcription styles are chosen through natural language instructions per request, eliminating the need for retraining.
The model achieves state-of-the-art speaker-attributed ASR accuracy in overlapping multi-speaker speech, outperforming specialized cascaded systems. Mean word error rates (WER) are 4.71% on 9 public English test sets and 4.62% on the multilingual leaderboard. Contextual transcription reduces named-entity recognition errors sixfold when given expected entity lists.
Efficient Inference and Training Strategies
For efficient inference, Nemotron 3.5 incorporates native Multi-Token Prediction (MTP) speculative decoding and an optional DFlash2 draft model. Using a single RTX PRO 6000 Blackwell GPU, the NVFP4 checkpoint achieves transcription speeds of approximately 2,238 seconds of audio per second.
Training utilizes a supervision-balanced mixture, coupling speaker activity from a frozen diarization model with ASR encoder representations. This enables speaker-attributed transcription through serialized output training without LLM architecture changes. The model undergoes supervised fine-tuning followed by a brief reinforcement learning stage to optimize transcription accuracy across multiple languages and tasks. Practical prompting and fine-tuning allow adaptation for new languages, domains, or tasks within 0.5 to 2.3 GPU hours on a single GPU.
Editor's take: Opportunities and Validation Approaches for Entrepreneurs
Nemotron 3.5 Transcribe offers a breakthrough integration of multilingual, speaker-attributed, and flexible instruction-driven transcription capabilities, making it especially suitable for startups focusing on voice interaction, meeting transcription, customer service, or content creation. Its open weights and fine-tuning toolkit lower barriers for developing customized transcription services.
Entrepreneurs can validate suitability by testing Nemotron 3.5 on their own or target market multilingual audio recordings, assessing speaker attribution accuracy, timestamp detail, and transcription quality following natural language instruction prompts. Comparing speed and accuracy against existing solutions will reveal integration opportunities and application fit. Its efficiency enables real-time or near-real-time services even with limited hardware resources, unlocking commercial potential.