If you are choosing an open-weight speech-to-text model this year, the Nemotron 3.5 Transcribe vs Whisper question comes down to what you need back from the audio. Do you want plain text, or text with speaker labels, translation and prompt-controlled formatting? Voxtral from Mistral AI is the third option, and its open-weight and hosted models are split across different products.
How this article was made: it is based on official announcements, documentation and the sources listed at the end, not on our own hands-on testing. See our editorial policy.
Table of Contents
Quick summary: Nemotron 3.5 Transcribe vs Whisper for each job
This is a summary of published specs, not a test result. Benchmark figures below come from the vendors or from leaderboards they cite, so treat them as vendor claims.
- Nemotron 3.5 Transcribe: best fit for recorded files where you want speaker labels, translation and custom formatting from a single model.
- Whisper: best fit if you need the broadest language list, a small footprint and a permissive MIT license.
- Voxtral: best fit for live, low-latency use through Voxtral Realtime. Its batch model, Voxtral Mini Transcribe V2, is not open-weight.
What each model is
Nemotron 3.5 Transcribe
Nvidia released Nemotron 3.5 Transcribe on Wednesday, October 28, as an open-weight speech model covering 40 languages. It couples a Transformer speech encoder and a frozen Nemotron 3 Diarization model with the Nemotron 3.5 Lightning 30B-A3B hybrid Mamba–Transformer mixture-of-experts LLM, for approximately 4.3B active parameters.
The output style is set per request. The transcription style is chosen with a natural language instruction for each request, without retraining. Nvidia says the model transcribes 40 languages, translates 31 of them to or from English, and identifies speakers in 13.
Whisper
Whisper is OpenAI’s long-standing open speech recognition model. Its code and model weights are released under the MIT License. The Hugging Face listing for large-v3-turbo shows 99 languages.
For speed-sensitive work, the repository describes the turbo model as an optimized version of large-v3 with faster transcription and minimal accuracy loss. The same README lists turbo at 809M parameters and roughly 6 GB of VRAM.
Voxtral
Voxtral is a family, and the open-weight status differs by model. For Voxtral Mini (3B) and Voxtral Small (24B), Mistral released the weights under the Apache 2.0 license. That was the July 2025 release.
The newer Voxtral Transcribe 2 split things differently. The family includes Voxtral Mini Transcribe V2 for batch transcription and Voxtral Realtime for live applications, and Voxtral Realtime is open-weights under Apache 2.0. The closed-weight model is called voxtral-mini-latest and is accessed through the Mistral API.
Comparison table for Nemotron 3.5 Transcribe vs Whisper (confirmed data only)
| Feature | Nemotron 3.5 Transcribe | Whisper (large-v3 / turbo) | Voxtral Realtime (open weights) |
|---|---|---|---|
| License | OpenMDW-1.1, per one third-party summary; check the model card | MIT | Apache 2.0 |
| Languages | 40 for transcription, 31 for translation | 99 (large-v3-turbo listing) | Not confirmed in the sources I read |
| Speaker labels | Yes, in 13 languages, up to 8 speakers | Not confirmed in the sources I read | Not confirmed for the open Realtime model |
| Translation | 31 languages to or from English | Not covered here | Not covered here |
| Prompt-controlled output | Yes, via text instruction | Not confirmed in the sources I read | Not confirmed in the sources I read |
| Streaming | No; it targets recordings, and a separate 0.6B streaming model exists | Not covered here | Yes, latency configurable down to sub-200ms |
| Size | About 30B total in the model name, about 4.3B active | Turbo: 809M parameters | 4B model, 8.87GB download |
Sources for the table: a separate Nemotron 3.5 ASR model of about 600M parameters is built for live audio, while Transcribe is the promptable file model for recordings up to 60 minutes. Voxtral Realtime is officially named Voxtral-Mini-4B-Realtime-2602 and is an 8.87GB download from Hugging Face.
Accuracy and speed claims
These numbers are vendor-reported or come from leaderboards the vendor cites, not independent tests. NVIDIA says it is the best open-weights model on the English and multilingual Open ASR Leaderboards. In its Hugging Face post, NVIDIA reports it ranked first among open-weight models with an average word error rate of 5.33% and an RTFx of 462.09.
On throughput, the company says the model processes 100 hours of audio in under three minutes on an RTX PRO 6000 Blackwell Server Edition at NVFP4 precision. That is a high-end data-center GPU. Expect lower numbers on other hardware.
On the diarization side, NVIDIA reports the lowest diarization error rate of the 14 systems tested on a 21-language benchmark track.
Voxtral’s older claims are also self-reported. For the original release, Mistral said Voxtral outperforms Whisper large-v3. That comparison predates Nemotron 3.5 Transcribe and says nothing about it directly. I found no independent head-to-head of all three models.
Key differences in Nemotron 3.5 Transcribe vs Whisper that affect your architecture
One model versus a pipeline
A common Whisper setup adds separate tools for speaker labels and alignment. Nemotron 3.5 Transcribe is designed to fold those jobs into one checkpoint. MarkTechPost says it can replace separate ASR, diarization and alignment tools. That can simplify deployment, but it also ties you to a much larger model.
Context and domain vocabulary
If your audio is full of product names or medical terms, this matters. Given a list of expected entities, NVIDIA says Nemotron 3.5 Transcribe reduces named-entity errors on ContextASR-Bench six-fold. Voxtral Mini Transcribe V2 offers a similar idea: context biasing lets you guide the model with custom vocabulary.
Hardware footprint
Whisper turbo is the lightest of the three at about 6 GB of VRAM. Nemotron is much heavier. Its model card lists a single H100 80 GB as the reference setup, and NVIDIA says it runs on one GPU from Jetson AGX Thor and DGX Spark up to Blackwell data-center cards. Budget for that before assuming it fits your current servers.
Limits to check
Audio, your prompt and the model’s response share a single 64,000-token context budget. That helps explain the roughly one-hour recording limit. Audio question answering was trained only in English, and the model card warns that language support differs by task. Nvidia’s own guidance is to spot-check names, speaker changes and answers against the original recording.
Which one to choose between Nemotron 3.5 Transcribe vs Whisper
- Meetings, calls and interviews with speaker attribution: start with Nemotron 3.5 Transcribe, provided you have the GPU capacity and your languages are among the 13 with speaker labels.
- Many languages, small hardware, or a permissive license: Whisper turbo is the safer baseline. The MIT License is the simplest to clear with legal teams.
- Live voice agents: Voxtral Realtime. It is available via API at $0.006 per minute and as open weights on Hugging Face. If you only want a hosted batch API, Simon Willison reported voxtral-mini-latest at $0.003 per minute, or $0.18 per hour. That is a hosted price, not self-hosting.
- Hosted instead of self-hosted: see our explainer on MAI-Transcribe-2 and our comparison of AI voice agent platforms.
How to evaluate Nemotron 3.5 Transcribe vs Whisper on your own audio
Public leaderboards use clean benchmark sets. Before committing, run a small sample of your own recordings through each candidate and compare word error rate, speaker errors and cost per audio hour. The right answer in the Nemotron 3.5 Transcribe vs Whisper decision depends on your accents, noise and domain terms. Public rankings cannot tell you that.
Conclusion
On paper, Nemotron 3.5 Transcribe is the most capable open-weight option for recorded audio that needs speaker labels, translation and prompt-controlled output. Whisper remains the lightweight, MIT-licensed, 99-language baseline, and Voxtral Realtime covers low-latency streaming under Apache 2.0. The Nemotron 3.5 Transcribe vs Whisper choice is therefore a trade between features and footprint. Validate it on your own data, because the accuracy figures here are vendor-reported.