Nemotron 3.5 Transcribe vs Whisper and Parakeet Compared

On October 28, 2026 NVIDIA published Nemotron 3.5 Transcribe, an open-weight speech model for recorded audio under the OpenMDW 1.1 licence. The Hugging Face blog announcing it describes a model that transcribes 40 languages, translates 31 of them to or from English, and attributes speech to up to eight speakers in 13 languages, on recordings up to an hour long. For a team that already runs Whisper large-v3 or Parakeet for batch transcription, the question is whether this replaces either, and the honest answer depends on which number you read and where it was measured.

This post works from three model cards and one leaderboard. The model cards supply licence, languages, parameters, hardware and the word error rates each vendor reports for its own model. The Open ASR Leaderboard, maintained by Hugging Face's audio team and read on November 5, 2026, supplies one comparable column for all three. The two kinds of figure are kept apart throughout, because the leaderboard's current default average excludes LibriSpeech and none of the model cards were measured on it. For the hosted services this post does not cover, our August comparison of GPT Transcribe, Whisper, Deepgram and Scribe is the companion piece.

NVIDIA's announcement is a Hugging Face blog post dated October 28, 2026, and its first sentence is the specification: an open-weight, promptable model that turns recorded speech into useful text. "Promptable" is the operative word. Instead of separate models for transcription, translation and diarization, one model takes audio plus a text instruction and returns text, which is why the same weights can label speakers in a meeting recording and translate a podcast. The blog is equally direct about the limit: "Transcribe is built for recorded audio, not live streaming."

The model card on Hugging Face fills in the numbers. The model has 34 billion parameters including the audio encoder, with about 4.3 billion active per token; weights are published in BF16 as the reference, with FP8 and NVFP4 variants. It handles recordings up to one hour and 65,536 tokens, and the card says it is not guaranteed to generalise beyond those limits. The blog text and the repository name say 30 billion total parameters while the card says 34 billion; this post uses the card's figure and notes the discrepancy rather than choosing silently. The training data is NVIDIA's Granary corpus, which the blog says now covers 463,062 hours across 48 languages in its second version.

The failure case to avoid is treating this as a Whisper drop-in. Whisper's receptive field is 30 seconds and it runs on a laptop; Nemotron's card lists an H100 for deployment and processes hour-long files in one pass. They are different shapes of tool, and the sections below are meant to make the shape visible before the error rates distract from it.

Each column below comes from that model's own card on Hugging Face, read November 5, 2026. Whisper large-v3 is OpenAI's largest Whisper checkpoint, an encoder-decoder transformer trained on weakly supervised data and released under Apache 2.0; its card lists 99 languages and 1,550 million parameters. Parakeet TDT 0.6B v3 is NVIDIA's 600-million-parameter transducer model, released August 14, 2025, extended from English to 25 European languages and licensed CC-BY-4.0. Where a card does not state a value, the cell says so.

The Open ASR Leaderboard is a Hugging Face Space run by the company's audio team, with the evaluation scripts in a public repository and a 2025 paper describing the method. It ranks models by average word error rate across a fixed set of English test datasets, lower being better, and reports an inverse real-time factor for open models that the maintainers can run. When read on November 5, 2026 the page stated it was last updated on November 2, 2026 and ranked 63 models. Its changelog records that the November 2 version removed the two LibriSpeech sets from the default datasets, so the average is now taken over nine sets: cleaned versions of AMI, Earnings22, GigaSpeech and VoxPopuli, SPGISpeech, URGENT24, Voice Arena Monsoon, and two private sets, one scripted and one conversational.

The rows that matter for this comparison are below. The eight entries above Nemotron, from Zoom, Otter, p99lab, Microsoft, Modulate, ElevenLabs, Reson8 and Sophea, all show their real-time factor as not available, which the board uses for systems its maintainers cannot run, so Nemotron 3.5 Transcribe is the first entry with published weights. The board lists four Parakeet TDT 0.6B entries: two v2 entries (rank 24 at 6.88% and rank 26 at 6.91%) and two v3 entries (rank 29 at 6.93% in a fast-gpu-asr build and rank 33 at 7.03%), so the range below mixes versions.

Each model's own documentation also quotes a leaderboard average, and none of them equals the figure above. NVIDIA's blog says Nemotron scores 5.33% average word error rate on the Open ASR Leaderboard, first among open-weight models. Nemotron's model card reports a 4.71% mean over nine datasets of its own listing, which includes both LibriSpeech sets and Monsoon English but not TEDLIUM. Parakeet's card reports 6.34% over an eight-set composition that includes TEDLIUM and both LibriSpeech sets. Whisper large-v3's card metadata reports 7.44% on the leaderboard as it stood whenever that card was last evaluated.

The reason is mechanical. The leaderboard's test-set composition has changed more than once; as read on November 5, 2026 it excludes LibriSpeech, and LibriSpeech is the easiest set in the group: Nemotron's card shows 1.09% on its clean split against 11.00% on URGENT24. Drop the easy sets from the average and every model's number rises. Each card was measured against whatever composition existed on its evaluation date, and none was re-measured on November 2. The three self-reported figures are therefore three different tests that happen to share a name.

The practical rule, and the one this post follows, is never to place a card figure beside a board figure. Compare cards with cards only where the dataset lists are identical, and otherwise compare on the board, where the maintainers ran every model on the same sets on the same date. The gap between the two kinds of number is also the gap between a launch post and a benchmark, which is why the same discipline appears in our benchmark methodology guide.

The three models span two orders of magnitude in size, and the hardware sentences on their cards reflect it. Nemotron's card lists one H100 80 GB for deployment, with a compatibility table for other GPUs; NVIDIA's blog names DGX Spark and Jetson AGX Thor for local use and Blackwell parts for the data centre. The headline speed claim is NVIDIA's own: 100 hours of audio in under three minutes on one RTX PRO 6000 Blackwell Server Edition GPU, measured on the forthcoming NVFP4 build with 30-second clips at high concurrency. That is a vendor throughput figure under favourable conditions and should be read as a ceiling. On the leaderboard, which measures on its own hardware, Nemotron's inverse real-time factor is 472.62, against 835.62 for the similarly accurate Qwen3-ASR-1.7B.

Parakeet's card says the model needs at least 2 GB of RAM to load, lists T4, L4, A10, A30, A100, L40, V100 and H100 as tested hardware, and transcribes up to 24 minutes of audio with full attention on an A100 80 GB. Whisper large-v3's card gives no hardware requirement; at 1.55 billion parameters it runs on consumer GPUs and, slowly, on CPUs, which is why it is still the default in so many pipelines. The failure case is choosing Nemotron for a workload that Parakeet already serves: a 600-million-parameter model that fits on a T4 and a 34-billion-parameter model that wants an H100 are not competing for the same slot unless the extra languages, speakers or translation are actually used.

Nemotron 3.5 Transcribe ships under the OpenMDW License Agreement version 1.1, the licence NVIDIA has used across its recent Nemotron releases. Whisper large-v3 is Apache 2.0, the most familiar of the three to a software team. Parakeet TDT 0.6B v3 is CC-BY-4.0, a Creative Commons licence whose governing-terms line the card states directly, and which carries an attribution requirement that a product shipping the model must honour.

None of the three restricts commercial use on its face, and this post does not offer legal advice on any of them. The practical point is that they are different documents with different attribution and notice duties, so a team swapping one model for another is also swapping licence obligations, and a compliance review done for Whisper does not transfer to Nemotron. The same is true of the usage caveats on the cards: Whisper's warns against transcribing recordings made without consent and against subjective classification of speakers, and Nemotron's states that speaker labels do not establish identity. Both belong in a deployment policy regardless of licence.

Strip the error rates away and the comparison is simpler than it looks. Whisper and Parakeet produce a transcript. Nemotron produces a transcript with speaker labels and, on request, a translation, from one set of weights. For a meeting-notes or call-review pipeline, speaker attribution is usually the feature that previously required a second model and a stitching step. NVIDIA's own blog makes the point by listing its earlier two-model system, Multitalker Parakeet plus Nemotron 3 Diarization, as the second-placed entry on the same diarization benchmark Nemotron 3.5 Transcribe now leads.

The diarization figures are NVIDIA's and should be read as such. The blog reports 32.4% concatenated-permutation word error rate on VoiceArena's Diarization-Bench, first of 14 systems on English full sessions, and an 18.26% diarization error rate on David AI's multilingual set against 18.38% for the previous Nemotron 3 Diarization model. The model card adds per-set multispeaker figures from 10.29% on AMI close-talk to 25.91% on NOTSOFAR-1 single-channel. A 32% figure on a hard benchmark is not a finished product; it is a reason to test on your own meeting audio, where microphone placement and cross-talk will decide whether the labels are usable.

Translation follows the same logic. Thirty-one languages to or from English is broader than Whisper's translate-to-English mode, and it is prompted rather than switched, so a single batch job can return the transcript and its English rendering together. Where a team needs a managed pipeline rather than weights, the hosted alternatives we covered in our look at Mistral's managed voice pipeline are the comparison to run. If you want help building the evaluation, with your own recordings, speakers and languages, before committing a GPU to any of these, it is part of what our AI transformation engagements cover.