NVIDIA folds transcription, speaker labeling and translation into one open-weight model: 40 languages, one-hour audio

One open-weight model now handles transcription, speaker diarization, translation and audio question answering, so a single recording no longer needs separate systems for each output. A transcript could previously capture every spoken word while still leaving unclear who said what, how to translate it, or how to format domain jargon, and multilingual calls, fast meetings and medical dictation each needed different output. The model ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for English and appears on Artificial Analysis's accuracy versus active-parameter frontier.

The output form follows a natural-language instruction, so one meeting recording can yield speaker labels, an English translation or a summary of decisions. This moves task selection from a fixed pipeline to the prompt layer, and the model card supplies recommended presets, a full prompt guide and a task-by-language support matrix. For speaker attribution it posts the lowest error rate on David AI's 21-language diarization leaderboard and ranks first in VoiceArena's 14-system comparison.

Prompt-level control extends into training, where developers adapt the model to the speech their applications encounter using Granary v2 multilingual data and NeMo Speech fine-tuning tools. Domain vocabulary and accents previously relied on context supplied in the prompt, and the medical case shows fine-tuning going beyond the best prompt. In a medical case study, fine-tuning cut word error rate on held-out speakers by 45.8% relative to the best prompt.

The demonstrated scope is the open-weight model's placement on public English transcription and speaker-attribution leaderboards plus the word-error-rate change on held-out speakers in one medical fine-tuning case; language support differs by task, and audio question answering is trained in English. The material states that quantized versions are coming soon for local runs on RTX Spark, DGX Spark or Jetson AGX Thor, or scaling on Blackwell data-center GPUs.

The supplied material is a summary, so the exact evaluation settings and comparison systems behind the leaderboard placements are not fully laid out; audio question answering is trained only in English and language support varies by task; the medical case reports a relative word-error-rate reduction against the best prompt without an absolute figure. The material also advises spot-checking names, speaker changes and answers against the recording for any speech model.