SOV-Transcribe-1 vs Nemotron 3.5 Transcribe: Fifteen Days Apart, and Only One of Them Is a Product

Open-weight speech recognition changed hands twice in a fortnight. NVIDIA published Nemotron 3.5 Transcribe on 28 October 2026; Salesforce published SOV-Transcribe-1 on 12 November. On Artificial Analysis' composite, the newer model wins: SOV-Transcribe-1 at 2.28% AA-WER against Nemotron 3.5 Transcribe at 2.81%. That is the whole of the sales pitch, and it is thinner than it looks. NVIDIA shipped speaker diarization, word timestamps, forty languages, one-hour recordings, three quantisation formats, a published technical report and a serving stack with a cookbook. Salesforce shipped the best number on the board and a model card that says the technical report is coming. If you are choosing a model to put into a product rather than a leaderboard, the accuracy gap is not the axis that matters.

Two architectures, two entirely different bets

Nemotron 3.5 Transcribe is a Mamba-2 and Transformer hybrid mixture-of-experts language model — Nemotron 3.5 Lightning 30B A3B — wearing a parallel expert encoder: a Transformer ASR encoder of roughly 600M parameters and a separate 99M-parameter speaker encoder, plus a CTC head for timestamp alignment. Thirty-four billion parameters total, 4.3B active including the audio encoder.

SOV-Transcribe-1 goes the other way. It takes an off-the-shelf Whisper large-v3 encoder, bolts on a projector, and hands the result to a Gemma 4 26B A4B decoder — 26.6B total, 4.6B active. Where NVIDIA built a speech front end and attached its own language model, Salesforce attached a general-purpose language model to a speech front end everybody already has. That is why the SOV checkpoint is 53 GB in BF16 and the Nemotron BF16 release needs an 80 GB H100: the Salesforce model carries less total weight, and it carries it in a format with no quantised variant published yet.

• Total parameters — SOV-Transcribe-1 26.6B versus Nemotron 3.5 Transcribe 34B, both counting the audio encoder.

• Active parameters — 4.6B versus 4.3B. Effectively a tie, which is the interesting part: Salesforce gets a lower error rate from a marginally larger active set and a substantially smaller total one.

• Audio limit — 30 minutes versus 1 hour, with a 65,536-token context on the NVIDIA side.

• Languages — 7 European versus 40, with code-switching supported across the Nemotron matrix.

• Precision — BF16 only (FP32 router) versus BF16, FP8 and NVFP4 checkpoints, the last of which runs on an RTX 5090.

• License — Apache 2.0 versus OpenMDW-1.1. Apache is the more permissive of the two by a distance, and for some legal teams that alone decides it.

One benchmark both vendors ran, with numbers that do not agree

This is the most useful comparison in the article, because it is the only place the two model cards can be put side by side on the same test. VoiceCodeBench measures structured-value transcription — email addresses, file paths, identifiers, account numbers — and reports two metrics: canonical token and entity match, which is how much of the target content was recovered, and task success rate, which is the share of recordings where every target value came out right.

NVIDIA's card publishes both: 88.53% CTEM and 59.33% TSR, plus a word error rate of 8.38% on the set. Salesforce's card publishes two figures for the same benchmark: 76.3% passing every target entity and 93.4% of entities recovered, described as first among open-weight systems and second of 23 overall. Those map onto TSR and CTEM respectively, and the mapping does not work. Salesforce claims 76.3% full-pass where NVIDIA reports 59.33% on the same metric, and claims 93.4% entity recovery where NVIDIA reports 88.53%.

There are innocent explanations. VoiceCodeBench is a private leaderboard run by a third party, scores are versioned over time, and the two vendors may have submitted different snapshots. There is also an uninnocent one: one of the two numbers is being described with more confidence than the benchmark's own reporting supports. What a reader can conclude is narrow and worth stating plainly — on the one structured-value benchmark both models were measured against, the Salesforce model appears to be meaningfully better at getting every value right, and the size of that advantage is contested by five to seventeen points depending on whose card you read.

What NVIDIA shipped that Salesforce did not

The gap here is not incremental. These are capabilities, not scores, and each one removes a piece of downstream work you would otherwise build.

• Speaker-attributed transcription — a full multi-speaker evaluation across AMI, CALLHOME, Mixer 6 and NOTSOFAR-1, with a mean cpWER of 15.80 and per-condition results down to individual distant-microphone cases. SOV-Transcribe-1 documents no diarization at all.

• Word timestamps — a CTC alignment head returning word text with start and end times, a diarization segment map and an alignment-error field. The Salesforce card returns a transcript string.

• Speech translation, language identification and audio question answering — separate tasks with their own benchmarks, including 100% FLEURS language ID on several languages. SOV-Transcribe-1 transcribes, in seven languages, and stops there.

• A technical report — published as a PDF, with the training-data composition itemised down to which corpora were licensed from whom. Salesforce links its report as "coming soon".

• A quantisation ladder and a serving cookbook — NVFP4 on consumer Blackwell, an SGLang-Omni path, GGML on native Windows. The Salesforce integration is a vLLM plugin pinned to 0.32.0 with upstream support described as work in progress.

The honest caveat on that list is that breadth and depth trade against each other. A model trained on 40 languages and seven task families spreads its capacity; one trained on seven European languages and transcription alone can spend all of it there. The 0.53-point composite gap is exactly what that trade looks like when it pays off.

The serving numbers, read carefully

Both vendors benchmark on the same card, which makes this comparison unusually clean — and unusually easy to get wrong.

• Salesforce, SOV-Transcribe-1 — up to 1,039 seconds of audio per second on one RTX PRO 6000 Blackwell Server Edition, with the DFlash draft model enabled, on a vLLM development build, at max_num_seqs=64.

• NVIDIA, Nemotron 3.5 Transcribe — up to 2,238 seconds of audio per second on the same card in NVFP4 with DFlash2 speculative decoding, returning a 30-second clip's transcript in 0.13 seconds.

The NVIDIA figure is more than double, and NVIDIA says why in its own card: those optimisations live in the Nemotron Transcribe NIM and are not in the open-source release, which the card describes as roughly two times slower and intended for model development. So the like-for-like comparison — open vLLM against open vLLM, on one card — is approximately 1,119 versus 1,039 audio-seconds per second. NVIDIA is still ahead, by about eight percent, and it is ahead while also being the model with speaker labels and timestamps in the output stream.

What that means for a buyer is that neither model is bought on throughput. Both saturate a modern card; both need a dedicated one. The decision is which output you need and which download you can run, and on the second question Nemotron wins outright, because an NVFP4 checkpoint runs on hardware that costs a fraction of a 96 GB server part while the Salesforce model has no quantised release to run.

Where SOV-Transcribe-1 is genuinely the better choice

Three cases, and the first is the obvious one. If English is the language and entity accuracy is the product, 2.28% against 2.81% on the composite is a real difference, and the Salesforce model's weakest surface — 4.06% on Earnings22-Cleaned — is still better than what the composite implies for the competition. On Contextual Earnings-22, SOV-Transcribe-1 reports keyword F1 rising from 71.9 to 96.8 with the clip's own hints and holding 94.0 when given the full hour's keyword list as distractors. NVIDIA reports its own Contextual Earnings22 numbers in word error rate rather than F1, from 12.54% with no context to 9.58% local, so the two are not directly comparable — but the shape of the result, that hint lists help and distractor-heavy hint lists still help, is the same on both.

The second case is licensing. Apache 2.0 against OpenMDW-1.1 is not a formality for a company whose legal review has opinions about derived works. If you intend to fine-tune and redistribute, or to embed the model in a product you ship, the Apache path is shorter.

The third is simplicity of scope. A model that does one thing returns one thing, and a pipeline that expects a transcript string is easier to reason about than one receiving a structured object with speaker tags, timestamps and an alignment-error field it may not need. Teams that want to own their diarization — because they already run a speaker-embedding model they trust — get a cleaner input from the narrower model.

Choosing between them without a leaderboard

If the audio is a meeting, a call with more than one voice, or anything where "who said it" is part of the deliverable, the answer is Nemotron 3.5 Transcribe and the accuracy question is moot — SOV-Transcribe-1 cannot produce the field. If the audio is non-English, the same applies, more sharply. If the recordings run past thirty minutes or need to run on hardware smaller than a data-centre card, NVIDIA again.

If the audio is English, single-speaker or already diarized upstream, under half an hour, and the deliverable is text with the right proper nouns in it, SOV-Transcribe-1 is the more accurate transcriber available today and the Apache licence removes a negotiation. That is a narrower set of workloads than the launch coverage implied, and it is a real one.

Mixed pipelines are the normal outcome here too, and they raise a practical question that neither model card addresses: what reads the transcript afterwards. OrcaRouter puts 200+ text, image and video models behind one key at provider list pricing with 0% markup, so the summariser or extractor downstream can be swapped or failed over without a second contract — useful when the transcript source is still being decided. Neither of these speech models is served through it, and we do not claim otherwise; the routing starts where the text does.

What has changed in the fortnight between these two releases is not which model sits on top of the open-weight table. It is that the table no longer has a clear answer, and the answer now depends on whether you are buying a transcriber or a transcript pipeline. NVIDIA sells the second. Salesforce published the first, and did it extremely well.