Recently, AI chip and technology giant NVIDIA has officially open-sourced its latest large-scale speech transcription model—NVIDIA Nemotron 3.5 Transcribe. As a cutting-edge open-source model tailored for professional voice-processing scenarios, it demonstrates industry-leading performance across core metrics including multilingual transcription accuracy, long-audio processing, and multi-speaker separation.
In terms of underlying architecture and core parameters, Nemotron 3.5 Transcribe adopts an advanced Mixture of Experts (MoE) model. With a total parameter count of roughly 34 billion, it activates only 4.3 billion parameters per forward pass, balancing strong inference capabilities with operational efficiency. In audio processing, the model can handle continuous recordings up to one hour in length, meeting the needs of various long-form audio tasks, meeting minutes, and interview transcriptions.
Regarding multilingual support and translation capabilities, this model exhibits exceptional global adaptability: it not only supports precise speech transcription in up to 40 different languages but also enables bidirectional translation between English and 31 of those languages, greatly expanding its applicability in cross-border communication and multilingual content processing. Additionally, addressing the longstanding challenge of multi-person conversations—a persistent pain point in speech recognition—the model can distinguish up to eight distinct speakers simultaneously and assign voices in real time.
In authoritative benchmark tests and evaluations, Nemotron 3.5 Transcribe has delivered outstanding results:
In Open ASR’s English-language evaluation, its average word error rate (WER) dropped to just 5.33%, placing it at the top among comparable open-weight models.
In Artificial Analysis’s non-streaming tests, its word error rate further decreased to 2.8%, tying for first place on the open-source leaderboard.
Overall, NVIDIA’s newly open-sourced Nemotron3.5Transcribe, with its massive parameter count, exceptional noise-robust separation capabilities, and outstanding multilingual performance, provides a powerful new foundation for developers worldwide and the broader speech technology ecosystem.