Nemotron Transcribe: Speech In, Shaped Text Out
Hand over your audio together with a written brief, and Nemotron transcribe returns the transcript in exactly the format you described
None

None

None

Long Story Video Skill

Turn your idea into 10s–10min video—AI writes scripts, prompts, and generates footage automatically.

Ads Video Skill

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

AI Video Prompt Generator

Feedback

freeTrialImage.bannerPity

freeTrialImage.upgradeUnlock

  • ✓freeTrialImage.benefitHd
  • ✓freeTrialImage.benefitWatermark
  • ✓freeTrialImage.benefitUnlimited

nemotron transcribe

Nemotron transcribe reads your written brief and shapes the transcript to match — speaker labels, word timings and 40 languages in one pass.

All Tools

Discover our comprehensive AI-powered animation toolkit

A Transcription Model That Reads Your Written Brief

Every reader wants a different transcript. With Nemotron transcribe you spell out the convention in plain words and the output obeys — no post-editing pass required.

  • Styles You Set with a Single Sentence
    Verbatim, written-form, contextual or speaker-attributed — the convention is chosen per request simply by what the instruction says.
  • A Speech Model Built in Three Layers
    A speech encoder, a frozen diarization model and a 30B-A3B hybrid MoE LLM running at roughly 4.3B active parameters.
  • Open Weights Plus a Tuning Toolkit
    OpenMDW-1.1 weights arrive as BF16, FP8, NVFP4, GGUF and DFlash2 builds, alongside the NeMo Speech toolkit for Nemotron transcribe.

Three Moves From Checkpoint to Labelled Transcript

Stand up a checkpoint, write the instruction, then decode and align — that is the whole path to an attributed transcript with Nemotron transcribe.

What Nemotron Transcribe Brings to Your Audio

Accuracy that tops the Open ASR board, attribution that survives crosstalk, and NIM inference quick enough for production — the strengths Nemotron transcribe delivers.

Top-Tier Accuracy on the Open ASR Leaderboard

A mean WER of 4.71% across the nine English Open ASR test sets, and 4.62% on the multilingual board.

Who Spoke, Even When Voices Overlap

From one-minute clips to hour-long meetings, it is the first SALM to outperform cascaded pipelines on speaker-attributed ASR.

Entity Lists That Cut Name Errors

Supply the entities you expect and named-entity mistakes drop six-fold on ContextASR-Bench, thanks to the parallel expert encoder.

Timestamps Tied to Each Speaker

A compact CTC adapter re-aligns the finished transcript with the audio, so long overlapping calls still return per-speaker word timings.

Fast Decoding and High Throughput

The model's MTP layer or a DFlash2 draft model accelerates decoding; the NVFP4 NIM handles 2,238 audio seconds per second on one RTX PRO 6000 Blackwell.

Adapt to a New Task 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.

FAQ

Questions About Nemotron Transcribe, Answered

Benchmarks, language coverage, speaker timestamps, licensing and checkpoints — the answers people ask for most about Nemotron transcribe.

1

What exactly is Nemotron 3.5 Transcribe?

A speech-augmented language model that couples a speech encoder, a frozen diarization model and a 30B-A3B hybrid MoE LLM in one stack.

2

Which transcript styles can I ask for?

Verbatim, written-form, contextual, speaker-attributed and rich transcription, plus translation and language identification — each one selected by an instruction.

3

How does it score on public benchmarks?

A 4.71% mean WER over the nine English leaderboard sets, 4.62% on the multilingual board, and six times fewer named-entity errors on ContextASR-Bench.

4

How long can a recording be, and how many languages are covered?

A single Nemotron transcribe checkpoint covers up to one hour of audio in 40 languages, code-switched speech included.

5

Does it show who spoke and when?

Yes. Overlapping multi-speaker audio is attributed per voice, and per-speaker word timestamps are returned whenever you request them.

6

What licence and checkpoints ship, and is fine-tuning possible?

OpenMDW-1.1 weights are released as BF16, FP8, NVFP4, GGUF and DFlash2 files, and NeMo Speech fine-tunes them in about 0.5–2.3 GPU-hours per task.

Put Nemotron Transcribe to Work on Your Audio

Choose a checkpoint, serve it through the NIM, and send your first instruction-driven request — styles, translation, language ID and speaker timestamps across 40 languages.