NVIDIA has open-sourced its speech transcription model, Nemotron 3.5 Transcribe, built on a Mixture-of-Experts (MoE) architecture with approximately 34 billion total parameters and 4.3 billion activated parameters. The model can process recordings up to one hour long, supports transcription in 40 languages and bidirectional translation between 31 languages and English, and can identify up to 8 speakers with real-time speaker attribution. In terms of performance, it achieves an average word error rate of 5.33% on the Open ASR English benchmark, ranking first among open-weight models; on the Artificial Analysis non-streaming test, it achieves a word error rate of 2.8%, tied for the top spot among open-source models.