Nemotron 3.5 Transcribe
A speech transcription and audio understanding model released by NVIDIA. It supports 40 languages and recordings up to one hour in length, and provides prompt-guided transcription, speaker diarization and audio question answering. Multiple precision weights are released openly.
Core Facts
Timeline (last 90 days)
NVIDIA released the Nemotron 3.5 Transcribe model, which uses a MoE architecture with 30B total parameters and activates only 4B parameters at runtime.
A single Nemotron 3.5 Transcribe model covers high-accuracy transcription of 40 languages, with a speaker diarization error rate as low as 1.8%.
NVIDIA released the 30-billion-parameter mixture-of-experts speech recognition model Nemotron 3.5 Transcribe, which activates only 4 billion parameters during inference and supports 40 languages and speaker recognition.
Nemotron 3.5 Transcribe has a total parameter count of 30B and about 4B activated parameters (precisely 4.3B).
The original material did not disclose the specific training data composition, license term details, or per-language WER breakdown of Nemotron 3.5 Transcribe, nor did it include direct comparison results with other open-source models.
Nemotron 3.5 Transcribe achieves an average word error rate (WER) of 4.71% on the English Open ASR Leaderboard.
Nemotron 3.5 Transcribe natively supports speaker labels and word timestamps.
NVIDIA recently released the Nemotron 3.5 Transcribe speech transcription model.
Nemotron 3.5 Transcribe's day-0 serving strategy essentially treats ecosystem collaboration as part of the product capability.
Nemotron 3.5 Transcribe is offered for inference through vLLM and NeMo Speech on the day of release.
5 more timeline events
All Facts (15)
NVIDIA recently released the Nemotron 3.5 Transcribe speech transcription model.
85%UnverifiedNVIDIA released the 30-billion-parameter mixture-of-experts speech recognition model Nemotron 3.5 Transcribe, which activates only 4 billion parameters during inference and supports 40 languages and speaker recognition.
60%UnverifiedA single Nemotron 3.5 Transcribe model covers high-accuracy transcription of 40 languages, with a speaker diarization error rate as low as 1.8%.
50%UnverifiedNemotron 3.5 Transcribe achieves an average word error rate (WER) of 4.71% on the English Open ASR Leaderboard.
50%UnverifiedNemotron 3.5 Transcribe natively supports speaker labels and word timestamps.
50%UnverifiedThe original material did not disclose the specific training data composition, license term details, or per-language WER breakdown of Nemotron 3.5 Transcribe, nor did it include direct comparison results with other open-source models.
50%UnverifiedNemotron 3.5 Transcribe has a total parameter count of 30B and about 4B activated parameters (precisely 4.3B).
50%UnverifiedNemotron 3.5 Transcribe's day-0 serving strategy essentially treats ecosystem collaboration as part of the product capability.
50%UnverifiedNemotron 3.5 Transcribe supports 40 languages.
50%UnverifiedNemotron 3.5 Transcribe is offered for inference through vLLM and NeMo Speech on the day of release.
50%UnverifiedNemotron 3.5 Transcribe has prompt-guided transcription, speaker diarization, and audio question-answering capabilities.
50%UnverifiedThe weights of Nemotron 3.5 Transcribe are released in three precisions: BF16, FP8, and NVFP4.
50%UnverifiedThe official tweet did not disclose the model parameter count of Nemotron 3.5 Transcribe, the recognition accuracy for each language, the audio question-answering evaluation results, or the degree of quality loss at NVFP4 precision.
50%UnverifiedA single recording for Nemotron 3.5 Transcribe can be up to one hour long.
50%UnverifiedNVIDIA released the Nemotron 3.5 Transcribe model, which uses a MoE architecture with 30B total parameters and activates only 4B parameters at runtime.
50%