A transcript is not always a neutral rendering of speech: a meeting record, a verbatim research transcript, and clean subtitles may need different wording. This model makes those choices promptable, bringing transcription style, translation, speaker attribution, and audio question answering into one checkpoint instead of requiring separate task-specific systems.
Key Capabilities
- Prompt-controlled transcripts: Request verbatim speech, cleaned-up wording, or domain vocabulary. This is useful when the desired transcript format changes by audience or workflow.
- Long, multilingual recordings: It supports 40 languages and recordings up to one hour, with a 65,536-token context. That makes it suitable for substantial media and meeting files, though it is an offline model rather than a real-time streaming service.
- Speaker and timing information: Optional speaker attribution and word-level timestamps can help with reviewing and editing conversations. Speaker labels distinguish voices within a recording; they do not identify real-world people.
- One model, several audio tasks: Alongside transcription, it supports speech translation, language identification, and audio question answering. Reported open-weight leaderboard means include 4.71% WER for English ASR and 4.62% for multilingual ASR.
Who It’s For—and the Trade-Off
A good fit for developers building transcription, meeting-processing, translation, or audio-review workflows who can run a large model on suitable NVIDIA hardware and want to adapt output through prompts or fine-tuning. Look elsewhere for low-resource, real-time streaming or lightweight deployment: the checkpoint has 34B total parameters and lists one H100 80 GB as its deployment target. Audio understanding and spoken-instruction results were evaluated only in English, and text-only reasoning is not its main strength.
