Nemotron 3.5 Transcribe vs Gemini 3.5 Transcribe: 2.8% on the Slide, 3.4% on the Board

Google's Gemini 3.5 Transcribe sits at rank nine on Artificial Analysis's non-streaming speech board with a 2.6% error rate. NVIDIA's Nemotron 3.5 Transcribe, released as open weights on 28 October 2026, sits at rank twenty with 3.4%. Four-tenths of a point separates them on the only public index that scores both, and that four-tenths is close to the least useful fact in this comparison. The two were not built for the same job: one is a managed endpoint that cleans up a transcript before you see it, the other is a checkpoint you download that keeps eight speakers straight while they talk over each other. The ranking does not flatter either of them for the thing it was actually designed to do.

It is worth flagging one discrepancy up front, because it recurs in every article written about this model this week. NVIDIA's launch material reports 2.8% AA-WER and a tie for first among open-weight systems. The live board, read on 29 October, carries 3.4% at rank twenty. Both numbers trace back to the same index, and they are not measurements of the same run — which is the first practical lesson of this pairing, and the reason the accuracy section below is written the way it is.

What NVIDIA shipped, and what Google has been running since August

Nemotron 3.5 Transcribe is not a transducer and not a plain encoder. It is a speech-augmented language model: a speech encoder feeds a frozen Nemotron 3 Diarization model, whose speaker-activity signal is injected into the ASR encoder through what NVIDIA calls a Parallel Expert Encoder, and the decoder is a hybrid Mamba–Transformer mixture-of-experts LLM with roughly 4 billion parameters active per token out of about 30 billion total. The company's own materials use both 30B-A3B and 30B-A4B labels for the language-model half, and the Hugging Face sidebar lists 34B once the audio stack is counted, so treat any single parameter figure as approximate. The weights are free under OpenMDW-1.1, which allows commercial use.

Gemini 3.5 Transcribe has been generally available since 26 August 2026 at $2.00 per million audio input tokens and $12.00 per million text output tokens, which Google translates to roughly $0.005 per minute, or about $0.30 per audio hour. It has a free tier inside quota, a separate live-streaming SKU at roughly $0.009 per minute, and no weights at all.

The number that moved between the launch packet and the morning after

Artificial Analysis's AA-WER is built on three datasets — AgentTalk at 50%, VoxPopuli-Cleaned at 25%, Earnings22-Cleaned at 25% — and the index is a single blended score, not a per-domain one. That matters because AgentTalk is voice-agent speech, which is short, clean and heavily represented in a vendor's training mix if the vendor is selling voice agents. A model that wins the blend can lose your archive.

Read the board as it stands, and the ranking order is: MAI Transcribe 2 at 2.0%, Voxtral Small Transcribe 3 at 2.1%, Gemini 3.5 Transcribe at 2.6%, GPT Transcribe at 3.3%, Nemotron 3.5 Transcribe at 3.4%. NVIDIA's 2.8% claim, if it had held, would have placed the new model level with Gemini 3.1 Pro Preview and Voxtral Small, and comfortably ahead of OpenAI's batch model. At 3.4% it is behind all of them and tied with three systems including a Gemini Flash-Lite variant it was never positioned against. A 0.6-point revision is not a rounding error on an index whose entire top ten spans 1.7% to 2.9%.

The honest reading is not that NVIDIA lied. AA re-measures when a model is served on a new endpoint or re-run on updated inputs, and the launch figure almost certainly reflects NVIDIA's own submission rather than the board's default host. But it is exactly the kind of gap that turns into a procurement mistake: a team that budgets against "first among open weights" and then measures 3.4% has bought a model on a slide.

What Google sells that NVIDIA does not

The feature that decides more transcription contracts than WER is what happens to the words after recognition, and Gemini 3.5 Transcribe is built around it. Its default verbatim mode keeps fillers, false starts and repetitions for compliance work; its smart mode strips disfluencies, resolves self-corrections and normalises spoken numbers and currency in a single pass. A custom vocabulary biases recognition toward up to 1,000 terms, with Google recommending about 100 well-chosen ones. Neither of those exists in Nemotron 3.5 Transcribe, which can be *prompted* toward an entity list and can be told to produce verbatim or written-form output, but which leaves number formatting and self-correction repair to whatever reads the transcript next.

Google also wins on breadth of named locales: 85+, with automatic detection and intra-sentence code-switching, against NVIDIA's 40 for transcription and 31 for translation. Named beats counted. If your queue includes a language that is on Google's list and not on NVIDIA's, there is no benchmark to argue about.

What NVIDIA sells that Google will not

Speaker attribution is the reverse case, and it is sharper than the marketing suggests. Gemini 3.5 Transcribe documents up to eight speakers, but Google's own materials flag attribution for three or more speakers as experimental, and its blog leans on "up to three". Nemotron 3.5 Transcribe labels up to eight speakers, does so in 13 languages, and returns speaker-attributed word-level timestamps through a post-decoding CTC adapter that aligns the finished transcript to the audio.

• Parameters — 30B total with roughly 4B active, hybrid Mamba–Transformer MoE vs an undisclosed proprietary speech stack.

• Licence — OpenMDW-1.1, commercial use, weights downloadable vs API-only with no redistribution rights.

• Price — no published list rate; cost is GPU time, and an NVFP4 build on one RTX PRO 6000 Blackwell is claimed to process up to 2,238 seconds of audio per second vs $0.005 per minute blended.

• Languages — 40 transcribed, 31 translated to or from English, 13 with speaker labels vs 85+ locales with code-switching.

• Speakers — 8, documented and benchmarked vs 8 documented, 3+ flagged experimental.

• Audio length — 1 hour in a single pass vs 1 hour, dropping to 30 minutes when diarization or timestamps are requested.

• Streaming — none; recorded audio only vs a separate live SKU at roughly $0.009 per minute.

• Post-processing — prompt-controlled verbatim or written form; no built-in cleanup vs smart mode that removes fillers and repairs self-corrections.

The diarization numbers only one side publishes

This is where the comparison stops being close and starts being lopsided, and also where the sourcing gets thin. NVIDIA reports 18.26% diarization error rate across 21 languages on the David AI benchmark, lowest of 14 systems evaluated, with the next entry being its own Nemotron 3 Diarization model at 18.38% and everything else between 22.85% and 37.55%. On VoiceArena's Diarization-Bench it reports 32.4% concatenated min-permutation WER, best of 14, with the runner-up at 34.2%. On overlapping-speech English sessions it reports dropping 8.32% of reference words against a median of roughly 31%.

Google publishes no diarization error rate for Gemini 3.5 Transcribe at all. Not a worse one — none. That asymmetry cuts both ways and should be stated plainly: NVIDIA's numbers are vendor-reported on NVIDIA's harness, and there is no third-party number to check them against because the independent diarization boards that exist are populated mostly by the systems NVIDIA itself entered. What is not in doubt is the documented ceiling: Gemini caps at eight speakers with experimental attribution above three, and drops your maximum audio length from an hour to thirty minutes the moment you ask for speaker labels and word timing together. If your material is a six-person board meeting, that is a product limitation rather than a quality question, and no WER comparison will surface it.

What an hour of audio actually costs on each

Work it on a thousand hours a month, a realistic load for a contact centre or a media archive. Gemini 3.5 Transcribe at $0.005 per minute is $300 per thousand hours. Nemotron 3.5 Transcribe has no list price because there is nothing to list: you buy the GPU. One RTX PRO 6000 Blackwell running the NVFP4 checkpoint is the configuration NVIDIA demonstrates, and the throughput claim — up to 2,238 seconds of audio per second — comes with a condition the Hugging Face post states and the headlines drop: it was measured on 30-second clips at high concurrency. Long meetings with overlapping speakers will not hit that number.

The crossover is therefore a utilisation question, not a price question. A single card that costs roughly the same as a few months of managed transcription only pays for itself if you keep it busy, and it never pays for itself against $300 a month of API calls on a small queue. The case for the checkpoint is not cheaper minutes at low volume. It is that the minutes stop being a line item at all, the audio never leaves your network, and fine-tuning is a single-GPU afternoon — NVIDIA reports 0.5 to 2.3 GPU-hours per task, and a medical-speech adaptation that took word error rate from 30.8% to 11.8% on 3,619 clips, with held-out speakers improving from 22.44% to 12.17%.

NVIDIA's own research page is the only place the throughput and fine-tuning claims appear together, and it is worth reading next to the numbers above: the 2,238 seconds per second figure assumes the NVFP4 checkpoint and the DFlash2 draft model on Blackwell hardware, and the 0.5-to-2.3 GPU-hour fine-tuning claim assumes a single GPU and a task small enough to fit one. Neither is a headline you can copy into a capacity plan without those qualifiers.

Whichever side you land on, the transcript is not the end of the pipeline. It gets summarised, classified, searched and routed, and that downstream half is where a single endpoint matters: OrcaRouter carries 200+ text and multimodal models behind one key, with provider list prices passed through at 0% markup, automatic failover and a routing DSL that can send meeting minutes to one summariser and compliance review to another without a second integration. Neither Gemini 3.5 Transcribe nor Nemotron 3.5 Transcribe is on that catalogue today — speech-to-text is still a separate contract wherever you buy it — so the point is narrower than it sounds: pick the transcriber on transcription, and let the layer above it stay portable.

Which one belongs in your stack

Choose Gemini 3.5 Transcribe if your problem is the text: disfluent interview audio, call-centre speech full of "no wait, I meant", a locale list you need to check by name, and a managed endpoint you can put behind a key this afternoon. It is the more accurate model on the public board today by a margin that is small but real, it fixes the transcript in flight, and it costs thirty cents an hour with a free tier to test against.

Choose Nemotron 3.5 Transcribe if your problem is the room. More than three speakers, overlapping speech, thirteen languages of speaker labels, word timing that survives an hour-long file, and an audio policy that forbids sending the recording to a third party — those are the conditions where the open checkpoint is not a compromise but the only option that meets the requirement. It is also the only one of the two you can fine-tune, which for accented or domain-specific audio is worth more than four-tenths of a point.

What would change this verdict: an independent diarization rerun that puts Gemini anywhere near NVIDIA's 18.26% DER, or an AA re-measurement that settles Nemotron between 2.8% and 3.4%. Until one of those lands, the accuracy argument is a vendor's, and the feature argument is the one you can verify yourself in an afternoon with your own worst recording.