Nvidia's New Transcription AI Takes Instructions

Nvidia's new open-weight transcription model is a top performer that can be guided with prompts to improve accuracy for specific tasks.

Nvidia has released Nemotron 3.5 Transcribe, a new multilingual model for converting speech to text. According to The New Stack, the model immediately established itself as a top performer. It currently ranks first on the Hugging Face Open ASR leaderboard for English, achieving an average word error rate of just 5.33%. The model also leads in two separate benchmarks for speaker attribution, which involves correctly identifying who is speaking. Significantly, Nvidia is releasing it as an open-weight model, meaning developers have broad access to its underlying architecture. This release signals Nvidia's growing ambition to compete directly in the foundational model space, complementing its dominant position in AI hardware with powerful, accessible software.

What makes Nemotron 3.5 Transcribe particularly noteworthy is its novel 'promptable' design. Traditional speech recognition systems typically function as a fixed tool: you provide an audio file and receive a text output, with little room for guidance. This new model changes that dynamic by allowing developers to provide text-based instructions, or prompts, alongside the audio. For example, a developer could prompt the model with a list of uncommon names, specialized technical terms, or company-specific acronyms that are likely to appear in the audio. This context guides the model's transcription process in real time, helping it make more accurate choices when faced with ambiguous sounds. This method of providing in-context guidance avoids the need for expensive and time-consuming fine-tuning for every specific domain.

This release fits into a broader industry trend of making AI models more controllable and adaptable. For years, developers faced a difficult choice in speech recognition: use a highly accurate but proprietary API from a major tech company, or opt for a more flexible but often less capable open-source model. Nvidia's strategy directly addresses this gap by offering a model that provides both state-of-the-art performance and the freedom of an open-weight license. The concept of prompting, which has become standard for large language models, is now proving its value in other domains like audio processing. By applying this technique to transcription, Nvidia is pushing the industry away from one-size-fits-all solutions and toward more dynamic, programmable AI systems.

The practical implications for developers and businesses are immediate and significant. The ability to easily customize a top-tier transcription model for specific contexts opens up a wide range of new applications. Teams can now build more reliable tools for medical dictation, legal transcription, or summarizing technical meetings without investing heavily in custom model training. This lowers the barrier to entry for startups and allows larger companies to enhance their products with more accurate audio features. Looking forward, the open-weight nature of Nemotron 3.5 Transcribe will likely encourage a wave of community-driven innovation. We can expect to see new applications, further fine-tuning for niche languages, and continued research that builds on this promptable architecture, putting pressure on closed-source providers to innovate.

For developers, this 'promptable' model shifts transcription from a generic utility to a programmable tool. The ability to provide real-time context without fine-tuning lowers the barrier for building sophisticated, domain-specific audio applications, from medical dictation to technical meeting summaries, using an open-weight model.

Companies can now achieve highly accurate, context-specific transcriptions without the significant cost and time of custom model training. This enables new product features and can reduce reliance on expensive, closed-source transcription APIs, potentially disrupting the market for specialized ASR services.