Eleven Realtime Translation
Translate live speech into spoken audio in more than 90 languages through an enterprise WebSocket API.
NVIDIA Nemotron 3.5 Transcribe
huggingface.co
Transcribe and analyze recorded audio with a downloadable model controlled through natural-language instructions.
Visit website
Pricing
Free
Available on
Self-hosted inference
Category
Ai Audio
What is NVIDIA Nemotron 3.5 Transcribe?
NVIDIA Nemotron 3.5 Transcribe is an open-weight speech-and-audio model for developers processing recorded audio. It accepts recordings, text instructions, and optional context, then produces text responses. It supports recordings up to one hour and a 65,536-token context, with advertised coverage of 40 languages.
Natural-language instructions select the desired transcription style without retraining. The model also supports speech translation, speaker attribution, and optional word timestamps. It suits teams building meeting processing, media transcription, customer-service analysis, and audio understanding systems.
Key features
Instruction-controlled transcripts
Choose verbatim wording, readable written form, contextual transcription, speaker-attributed output, or rich transcription through each request's instructions.
Vocabulary and context guidance
Provide contextual information to guide recognition of domain-specific vocabulary and preferred transcript conventions.
Speaker attribution and alignment
Distinguish speakers within a recording and optionally request speaker-attributed word timestamps after transcript decoding.
Multilingual speech tasks
Support includes transcription, speech translation, language identification, and code-switching, with task availability varying across the language matrix.
Audio question answering
Answer questions about recordings and follow spoken instructions. These capabilities have been evaluated only in English.
Who is it for?
Meeting transcript pipelines
Application developers can prepare speaker-attributed transcripts for recorded meetings.
Specialist vocabulary processing
Domain teams can supply terminology context when processing technical interviews or service calls.
Multilingual media processing
Media developers can evaluate transcription and translation for the languages and tasks their catalog requires.
How to use NVIDIA Nemotron 3.5 Transcribe
- Check the model card's language matrix and hardware requirements for your intended task.
- Download a suitable checkpoint and follow the documented vLLM and NeMo Speech serving setup.
- Prepare recorded audio as 16 kHz mono and provide a text instruction describing the requested output.
- Submit the recording and optional vocabulary context; inspect the transcript and any requested alignment output.
NVIDIA Nemotron 3.5 Transcribe pricing
Free
Downloadable model materials are licensed free of charge under OpenMDW-1.1, subject to its terms. Hardware and hosting costs are separate.
Check current pricing ↗What to know before you use it
- This is an offline model, not a real-time streaming recognizer. Its internal 30-second audio chunks do not provide streaming recognition.
- Check the model card's language matrix for each task. Translation and multispeaker support vary by language; Norwegian is labeled supported but has no task checks in the matrix. Audio question answering and spoken-instruction following were evaluated only in English.
- Serving requires substantial GPU resources. The documented BF16 profile uses one H100 80 GB; NVFP4 requires Blackwell hardware. Serving profiles do not establish fine-tuning memory needs.
- Timestamp requests require one recording, non-streaming output, and one completion. DFlash2 serving cannot currently be combined with CTC timestamps; applications should inspect alignment errors.
Frequently asked questions
Does it support live transcription?
No. It processes recorded audio offline, including recordings up to one hour.
Can instructions change the transcript format?
Yes. Instructions can select formats such as verbatim, written-form, contextual, and speaker-attributed transcription without retraining.
Can it answer audio questions in every supported language?
Audio question answering and spoken-instruction following were evaluated only in English. Check the language matrix separately for transcription and translation support.