None
None
Long Story Video Skill
Turn your idea into 10s–10min video—AI writes scripts, prompts, and generates footage automatically.
Ads Video Skill
Generate professional ads and sales videos—AI auto-generates scripts, prompts, and footage.
3D Science Explainer Video Skill
Convert scientific concepts into stunning 3D explain animations
Feedback
freeTrialImage.bannerPity
freeTrialImage.upgradeUnlock
- ✓freeTrialImage.benefitHd
- ✓freeTrialImage.benefitWatermark
- ✓freeTrialImage.benefitUnlimited
nemotron transcribe
Prompt-driven speech recognition: describe the transcript format you want and Nemotron Transcribe applies it across 40 languages and hour-long audio.
All Tools
Discover our comprehensive AI-powered animation toolkit
Speech Recognition That Takes Direction
Every reader wants a different transcript. With Nemotron Transcribe you describe the convention in ordinary language, and the model applies it to the whole recording.
- Request Any Transcript ConventionAsk for verbatim, cleaned-up, context-aware or speaker-labelled output — the style is set per request rather than baked into the model.
- Speech-Aware LLM BackboneA speech encoder, a frozen diarization module and a 30B-A3B hybrid MoE language model with roughly 4.3B active parameters work as one.
- Open Weights Plus Fine-Tuning KitDownload OpenMDW-1.1 checkpoints in BF16, FP8, NVFP4, GGUF or DFlash2, then adapt them with the NeMo Speech toolkit.
Running Nemotron Transcribe in Three Moves
Deploy a checkpoint, phrase your instruction, then decode and align — that is the whole path to a speaker-attributed transcript.
What Nemotron Transcribe Delivers
Top-tier accuracy, per-speaker attribution on overlapping talk and fast NIM serving — the capabilities this model brings to production work.
Leaderboard-Topping Accuracy
A mean WER of 4.71% across the nine English Open ASR test sets, and 4.62% on the multilingual leaderboard.
Who-Spoke-When on Overlapping Audio
From one-minute clips to full meetings, it is the first SALM to outperform cascaded pipelines on attributed ASR.
Entity Lists Steer Contextual Accuracy
Supplying expected entities cuts named-entity mistakes six-fold on ContextASR-Bench, thanks to the parallel expert encoder.
Word Timings Tagged by Speaker
A compact CTC adapter maps the finished transcript back onto the audio, so even long overlapping calls keep per-speaker timings.
Fast Decoding, High Throughput
The MTP layer or a DFlash2 draft accelerates decoding; the NVFP4 NIM processes 2,238 audio seconds per second on a single RTX PRO 6000 Blackwell.
New Tasks on a Single GPU
A supervision-balanced mixture lets one GPU absorb new languages and domains in roughly 0.5–2.3 GPU-hours per task.
Nemotron Transcribe — Questions Answered
Straight answers on benchmark scores, language coverage, timestamps, licensing and the checkpoints that are available to download.
What exactly is Nemotron 3.5 Transcribe?
A speech-augmented language model that combines a speech encoder, a frozen diarization model and a 30B-A3B hybrid MoE LLM.
Which output styles can I ask for?
Verbatim, written-form, contextual, speaker-attributed and rich transcription, plus translation and language ID — each one selected by instruction.
How does it score on public benchmarks?
4.71% mean WER across the nine English leaderboard sets, 4.62% multilingual, and six times fewer named-entity errors on ContextASR-Bench.
What recording length and language coverage should I expect?
A single checkpoint handles audio of up to one hour in 40 languages, with code-switched speech included.
Does it identify who is speaking and when?
Yes. Overlapping multi-speaker audio is attributed per speaker, and per-speaker word timestamps are returned on request.
What is the license, which checkpoints ship, and can I fine-tune?
OpenMDW-1.1 weights arrive as BF16, FP8, NVFP4, GGUF and DFlash2 files, and NeMo Speech fine-tunes them in about 0.5–2.3 GPU-hours per task.
Try Nemotron Transcribe on Your Own Audio
Choose a checkpoint, serve it through the Nemotron Transcribe NIM and send an instruction-driven request — formatting styles, translation, language ID and speaker timestamps across 40 languages.