Recently, NVIDIA, the AI chip and technology giant, officially open-sourced a new large speech transcription model — NVIDIA Nemotron 3.5 Transcribe. As a cutting-edge open-source model built for professional speech processing scenarios, it demonstrates industry-leading capabilities on core metrics such as multilingual transcription accuracy, long-audio processing, and multi-speaker separation.

In terms of underlying architecture and core parameters, Nemotron 3.5 Transcribe uses an advanced Mixture-of-Experts (MoE) architecture. It has about 34 billion total parameters, while 4.3 billion parameters are activated per forward pass, ensuring strong inference capability while also maintaining runtime efficiency. For audio processing, the model can handle continuous recording files up to 1 hour long in a single pass, meeting transcription needs for long audio, meeting minutes, and interview programs.

In multilingual support and translation, the model shows strong global adaptability: it not only fully supports accurate speech transcription in as many as 40 different languages, but also supports bidirectional translation between 31 of those languages and English, greatly broadening its use cases in cross-border communication and multilingual content processing. In addition, to address multi-person conversations, a long-standing pain point in speech recognition, the model can distinguish up to 8 different speakers at the same time and complete speaker attribution in real time.

In authoritative benchmarks and evaluations, Nemotron 3.5 Transcribe also performs impressively:

In the Open ASR English evaluation, its average word error rate (WER) is as low as 5.33%, ranking first among comparable open-weight models.

In Artificial Analysis's non-streaming test, its word error rate drops further to 2.8%, tying for first place on the open-source leaderboard.

Overall, with its large parameter scale, outstanding interference-resistant separation capability, and excellent multilingual performance, NVIDIA's newly open-sourced Nemotron 3.5 Transcribe provides a powerful new foundation for global developers and the speech technology ecosystem.