NVIDIA-Nemotron-3.5-Transcribe-30B-A4B
Nemotron 3.5 Transcribe is an open-weight 30B-A4B speech-augmented language model from NVIDIA for offline transcription in 40 languages. It handles recordings up to one hour and adds speech translation, contextual and speaker-attributed transcription, word timestamps and audio question answering, leading open-weight models on English, long-form and multilingual ASR.
Nemotron audio transcription demo
Input
Audio preview stays in your browser. A paid plan is required before transcription can enter the queue.
Transcription settings
Output
Upload a recording, choose settings, then start the model.
Model informationEverything behind the transcription
Speech intelligence built for real recordings
Nemotron 3.5 Transcribe is an open-weight, offline speech and audio model from NVIDIA. It turns recordings, together with text instructions and optional context, into text: multilingual transcription in 40 languages, speech translation, contextual transcription with domain vocabulary, speaker-attributed transcription, rich transcription, word timestamps, language identification, and audio question answering. It handles recordings up to one hour with a 65,536-token context.
Model summary
| Parameters | 34B total, ~4.3B active (including the audio encoder) |
|---|---|
| Architecture | Speech-Augmented Language Model: Nemotron 3.5 Lightning hybrid Mamba-2–Transformer MoE backbone with a Transformer ASR encoder, a Nemotron-3-Diarization speaker encoder and a CTC timestamp head |
| Precision | BF16 |
| Context | 65,536 tokens; recordings up to 1 hour |
| Languages | 40 |
| Speculative decoding | Native Multi-Token Prediction (MTP) |
| License | OpenMDW-1.1 |
Results at a glance
Error rates in percent, lower is better (NVIDIA, October 2026).
| Capability | Benchmark | Nemotron 3.5 Transcribe | Next best open-weight |
|---|---|---|---|
| English ASR | Open ASR Leaderboard, 9 sets (mean WER) | 4.71 | Qwen3-ASR-1.7B 4.94 |
| Long-form ASR | Open ASR Leaderboard, long-form (mean WER) | 10.98 | Parakeet-TDT-0.6B-v3 13.37 |
| Multilingual ASR | Multilingual Open ASR Leaderboard, 7 languages (mean WER) | 4.62 | Qwen3-ASR-1.7B 6.49 |
| Multi-speaker ASR | AMI, CALLHOME, Mixer 6, NOTSOFAR-1 (mean cpWER) | 15.80 | DiCoW v3.3 + DiariZen 16.83 |
Prompting
Pick the transcription style with the instruction; keep the system prompt empty for transcription and translation.
| What you need | Instruction |
|---|---|
| Verbatim words | Audio transcription. Language: as spoken. Wording: verbatim, including fillers and repetitions. Formatting: punctuation and capitalization. Output: transcript only. |
| Readable text | Audio transcription. Language: as spoken. Wording: omit fillers and accidental repetitions; keep the remaining spoken wording unchanged. Formatting: punctuation and capitalization. Output: transcript only. |
| Domain vocabulary | Transcribe with punctuation and capitalization. Context terms: <names and terms>. |
| Who said what | Transcribe verbatim, including fillers. Write numbers as spoken, without punctuation or capitalization. Start every speaker turn with <spk:0>, <spk:1>, etc. Keep one stable tag per speaker. Output only the speaker-tagged transcript. |
| Translation | Please translate the given speech to <target language>. |
Use temperature 0 with reasoning off for transcription, translation and language identification.
This is an offline model: it is not a real-time streaming ASR service. For streaming transcription, see Nemotron 3.5 ASR Streaming Multilingual.
Ethical considerations
Speaker labels distinguish speakers within a recording; they do not establish real-world identity. NVIDIA believes Trustworthy AI is a shared responsibility. Developers should make sure the model meets the requirements of their use case and addresses unforeseen product misuse.
Use of this model is governed by the OpenMDW License Agreement, version 1.1.
Turn every conversation into useful text.
Explore Nemotron 3.5 Transcribe with a plan built for long-form, multilingual audio.