SOV-Transcribe-1 vs Whisper Large v3 Turbo: What Two Years and 2.3 Points of WER Are Worth

Whisper Large v3 Turbo is not a competitor to SOV-Transcribe-1 in any commercial sense. It is the incumbent — an 809M-parameter, MIT-licensed checkpoint published in October 2024 that quietly sits underneath a large share of every transcription product shipped since, and which still appears on hosted endpoints at $0.67 per thousand minutes. SOV-Transcribe-1 is the challenger, published by Salesforce on 12 November 2026, scoring 2.28% against Whisper Large v3 Turbo's 4.62% on Artificial Analysis' composite. So the question is not which model is better. It is whether halving the error rate is worth a 53 GB download, a 96 GB GPU and a serving stack, when the alternative fits in a laptop's memory and costs pennies per hour of audio.

What two years of progress actually bought

Whisper Large v3 Turbo is a pruned Whisper large-v3 with the decoder cut from 32 layers to 4. OpenAI's own model card is disarmingly direct about it: "the exact same model, except that the number of decoding layers have reduced from 32 to 4. As a result, the model is way faster, at the expense of a minor quality degradation." That sentence describes a compression decision, not an architecture.

SOV-Transcribe-1 is the other kind of change. Salesforce kept a Whisper large-v3 encoder — the same front end — and replaced the four-layer decoder with a Gemma 4 26B A4B mixture-of-experts language model behind a projector. The audio perception is inherited from Whisper. The thing that turns audio into words is a general-purpose language model thirty times the size of the entire Whisper turbo checkpoint.

• Parameters — Whisper Large v3 Turbo 809M versus SOV-Transcribe-1 26.6B total, 4.6B active.

• Decoder — 4 Transformer layers versus a 26B mixture-of-experts with 4.6B active per token.

• Checkpoint size — under 2 GB versus roughly 53 GB in BF16.

• Languages — roughly a hundred versus seven. This one runs the other way, and it runs hard.

• Timestamps — sentence- and word-level output built into the decoding loop versus none documented.

• License — MIT versus Apache 2.0. Both permissive; MIT is marginally simpler.

• Hardware — CPU or a small GPU versus an RTX PRO 6000-class card with 96 GB for the draft-enabled configuration.

That last line is the whole argument in miniature. One model runs anywhere; the other runs on a server part that costs more than most transcription projects' entire annual cloud bill.

The error arithmetic nobody does

Word error rate is an abstraction until you convert it into corrections. Take a corpus of a thousand minutes of English speech at a conversational 150 words per minute — 150,000 words. At 4.62%, Whisper Large v3 Turbo produces roughly 6,900 wrong, missing or inserted words. At 2.28%, SOV-Transcribe-1 produces roughly 3,400. The difference is about 3,500 words that a human does not have to find and fix.

Now price the fixing. A careful editor reconciles transcript against audio at perhaps 200 to 400 words per minute, so 3,500 fewer errors is somewhere between nine and eighteen minutes of human review saved per thousand minutes of audio. At a loaded rate of $30 an hour that is $4.50 to $9.00 — against a transcription cost, on the hosted Whisper side, of $0.67.

That comparison is the reason most teams never migrate. The human cost of errors dominates the machine cost of transcription by an order of magnitude, so a model that halves errors should be worth ten times the price and in practice is competing against a price that is already negligible. The exception is any pipeline where nobody reviews the output at all — automated indexing, search, analytics, summarisation feeding — because there the errors do not get fixed, they propagate into whatever reads the transcript next, and the second-order cost is the one that bites.

The index both scores sit on is itself a blend — half synthetic agent conversation, a quarter read parliamentary speech, a quarter real earnings calls — and the two models behave very differently on each third. Whisper Large v3 Turbo's components are 3.52%, 3.97% and 7.49% respectively; SOV-Transcribe-1's are 1.79%, 1.50% and 4.06%. The gap is widest on read parliamentary speech — 1.50% against 3.97%, a factor of 2.6 — and narrowest on earnings calls, where both models struggle and the gap is 4.06% against 7.49%. If your audio is clean read speech, the upgrade is dramatic. If it is messy multi-speaker financial audio, both models are bad and the newer one is merely less bad.

Why Whisper still wins the invoice

Start with the hosted comparison, because it is the one most teams actually face. Whisper Large v3 Turbo is available at $0.67 per thousand minutes from at least one hosting provider on the Artificial Analysis board, at a measured speed factor of 339.5 — that is 339.5 seconds of audio per second of wall clock. SOV-Transcribe-1 has no hosted per-minute price at all. Salesforce published weights, not an endpoint, and nothing on the board carries a per-1,000-minute figure for it.

So the comparison becomes dollars against GPU hours, and the model card gives unusually complete numbers for it. With native vLLM support and the DFlash draft model, one RTX PRO 6000 Blackwell Server Edition transcribes up to 1,039 seconds of audio per second. Run that card flat out for an hour and it processes 3.74 million seconds of audio — 62,340 minutes. At a couple of dollars an hour for the rental, that is on the order of three cents per thousand minutes of audio, twenty times cheaper than the cheapest hosted Whisper endpoint on the board.

That number is real and it is also the most misleading number in this article, because it assumes the card is never idle. Speech transcription is bursty: a backlog arrives overnight, gets cleared, and then the GPU sits at zero utilisation for twenty hours. At 1,039 audio-seconds per second, a single card clears a thousand hours of audio in under an hour. Almost nobody has a thousand hours of audio a day. The honest conclusion is that self-hosting either model is an economics decision about utilisation, not about size — and that once you have decided to run a card at all, the marginal cost of the better model is close to zero.

Three things Whisper does that the new model cannot

Language coverage is the first, and it is not close. Whisper Large v3 Turbo's model card lists close to a hundred languages, from English and Mandarin down through Welsh, Maltese, Hawaiian, Lao and Sundanese. SOV-Transcribe-1 covers English, French, German, Spanish, Italian, Portuguese and Dutch. For any team whose audio is not European, this comparison has no second act — Whisper is not the budget option, it is the only option between the two.

Timestamps are the second. Whisper's decoding loop predicts them natively, and the Transformers pipeline exposes them at sentence or word granularity through a single argument; that is how subtitles, chapter markers and compliance search have been built on Whisper for two years. SOV-Transcribe-1's card documents transcription, style control and context biasing, and nothing about timestamps. Recovering them downstream means running a forced aligner over the output, which is another model, another dependency and another source of error.

Speech translation is the third. Whisper can take audio in any of its source languages and emit English text directly, as a decoding task rather than a second pipeline stage. SOV-Transcribe-1 transcribes in the language spoken. If you need subtitles for a Japanese documentary, Whisper does it in one call and the Salesforce model cannot do it at all.

Whisper's documented weaknesses deserve equal billing, since they are the strongest argument for moving. OpenAI's own card warns that the model hallucinates — producing text that was not spoken — because it was trained on weakly supervised, noisy data, and that this is worse in low-resource languages. It also notes that the sequence-to-sequence architecture is prone to repetition, only partly mitigated by beam search and temperature fallback. Those are exactly the failure modes that make an unreviewed pipeline dangerous, and they are what two years of architecture work was aimed at.

What a real migration looks like

Almost nobody replaces Whisper with SOV-Transcribe-1. They route between them, and the routing rule writes itself from the specs: if the audio is English or one of the six other covered languages, single-speaker, under thirty minutes and entity-heavy, send it to the open model with a hint list; otherwise leave it on Whisper. Salesforce's Contextual Earnings-22 result — keyword F1 rising from 71.9 with no hints to 96.8 with the clip's own keywords, and holding 94.0 when handed the whole hour's keyword list as distractors — is the specific capability that makes the split worthwhile, because it is the one thing Whisper's prefix-prompt biasing does not do nearly as well.

The cost of that split is duplicate infrastructure and two output formats: SOV-Transcribe-1 returns a transcript string, Whisper returns text plus timestamp chunks, and everything downstream has to accept both. In practice most teams normalise to their own transcript schema at the boundary and stop caring which model produced it.

What happens after the transcript is the part that scales with volume, and it is where a router belongs rather than a second speech vendor: OrcaRouter puts 200+ text, image and video models behind one key at provider list pricing with 0% markup, so the summariser, indexer or classifier that consumes either transcript can be swapped or failed over without a new contract — and a vendor price cut on that model is live here the same day. Neither Whisper Large v3 Turbo nor SOV-Transcribe-1 is served through it; both are weights you run yourself, and the routing starts with the text they emit.

What would make this comparison obsolete

Two things, and both are plausible within a release cycle. The first is a quantised SOV checkpoint. Whisper Large v3 Turbo's practical advantage is that it fits anywhere; a 4-bit build of a 26.6B model would still want a serious GPU, but an FP8 or NVFP4 release — the kind NVIDIA shipped on day one for Nemotron 3.5 Transcribe — would move the Salesforce model from "data-centre part" to "workstation part" and collapse a large part of the deployment objection. Salesforce has published an MLX 4-bit build already, which suggests the direction.

The second is timestamps and diarization arriving in a point release. Both are conspicuously absent, both are table stakes for meeting and media products, and both are exactly the kind of thing a v0.2 model card adds. If they land, the only remaining reason to run Whisper is the hundred languages — and for a large share of teams, that is the only reason they run it now.

Until then the split is stable and easy to justify: Whisper Large v3 Turbo remains the correct default for anything multilingual, timestamped, long-tail or CPU-bound, and SOV-Transcribe-1 is the specialist you reach for when the audio is English, the vocabulary is known and nobody is going to read the output before a machine does.