SOV-Transcribe-1 Is the Most Accurate Open Speech Model. It Is Also the Least Convenient.

Salesforce's SOV-Transcribe-1 is the best open-weights speech recogniser anyone has published. On Artificial Analysis' non-streaming board it scores 2.28% AA-WER, which is lower than every proprietary model except a handful and lower than every downloadable one by a clear margin — Nemotron 3.5 Transcribe sits at 2.81% and the long-serving Whisper Large v3 Turbo at 4.62%. The catch is written into the same model card: SOV-Transcribe-1 is non-streaming, covers seven languages, ships as a 53 GB BF16 checkpoint, and is meant to be served by you on your own GPU. It was released on 12 November 2026 under Apache 2.0, and the accuracy is not the interesting part of the story. The interesting part is what it costs to turn that accuracy into a working pipeline.

What Salesforce actually put on Hugging Face

The repository is salesforce-sovereign-intelligence/SOV-Transcribe-1, created on 10 November and touched again on the 11th; the launch post and the Artificial Analysis placement both landed on the 12th. It is Salesforce's first open-weight speech-augmented language model, which is a specific architecture claim rather than marketing: an audio encoder feeds a projector, and a large mixture-of-experts decoder writes the transcript.

• Size — 26.6B parameters in total, 4.6B active per token, and both figures include the audio encoder. That pairing is what the launch leans on, because it puts the model on the accuracy-per-active-parameter frontier rather than the accuracy-per-total-parameter one.

• Architecture — a Whisper large-v3 audio encoder, a small projector that stacks four encoder frames into one embedding, and a Gemma 4 26B A4B mixture-of-experts text decoder. The audio path is inherited, the language path is Google's, and the contribution is the adaption between them.

• Precision — BF16 weights with the MoE router kept in FP32. The checkpoint is sixteen safetensors shards totalling roughly 53 GB, so "open weights" here means a download that will not fit on a single 48 GB card in BF16.

• Audio — 16 kHz mono input, up to 30 minutes per recording as tested. Longer files are the caller's problem: the recommended workaround is to split at speech boundaries and concatenate.

• Languages — English, French, German, Spanish, Italian, Portuguese and Dutch. Seven, all European, English dominant in every benchmark Salesforce published.

• Control — instructions in the user message. Verbatim or cleaned-up output, punctuation and capitalisation, digits versus spelled-out numbers, and optional context biasing with a keyword list or a sentence describing the domain.

• Frameworks — a vLLM plugin pinned to vLLM 0.32.0, or Transformers 5.19 through 5.22 with the repository's own modelling code. Native upstream support is described as work in progress, not shipped.

• Speculative decoding — a separate DFlash draft model, SOV-Transcribe-1-DFlash, proposes fifteen tokens at a time. Salesforce measures a 2–3x decode speedup for single-request serving, with accuracy "within run-to-run variation".

Where the 2.28% comes from

Artificial Analysis does not publish a single WER. Its AA-WER v2 index is a weighted composite of three sets: AA-AgentTalk at 50%, VoxPopuli-Cleaned-AA at 25% and Earnings22-Cleaned-AA at 25%. That weighting matters, because the three components disagree about this model by a factor of nearly three.

• AA-AgentTalk — 1.79% word error rate, the synthetic conversational set carrying half the index weight.

• VoxPopuli-Cleaned — 1.50%, read speech from European parliament recordings. This is the model's best surface and the one closest to its seven-language training mix.

• Earnings22-Cleaned — 4.06%, real earnings calls with names, tickers and numbers. This is the model's worst surface and the one closest to actual enterprise audio.

The ranking claim needs two numbers, not one. Artificial Analysis' own summary of the open-weight field is that SOV-Transcribe-1 leads it, with Mistral's Voxtral Small at 2.77% second and NVIDIA's Nemotron 3.5 Transcribe at 2.81% third — sixteen open-weight entries among 65 evaluated. Across the whole board, Artificial Analysis announced the model as seventh of 44 overall on 12 November. The board we read on 13 November places SOV-Transcribe-1 eighth, behind seven proprietary and newer systems including Alibaba's Fun-Realtime-ASR-preview, ElevenLabs Scribe v3 and Microsoft's MAI-Transcribe-2. Both statements are true on their own date; only one of them is true today.

The vendor's own numbers, and what they are worth

Everything in this section is Salesforce-reported. None of it has been reproduced outside the company, and two of the benchmarks are not public, which means nobody outside can reproduce them even in principle.

• Entity accuracy — on VoiceCodeBench, a private leaderboard run by BeSimple, SOV-Transcribe-1 passes 76.3% of recordings outright (every target entity correct) and recovers 93.4% of entities. Salesforce reports that as first among open-weight systems and second of 23 overall. Because the set is private, the only evidence available is the vendor's summary.

• Accents — 8.21% average WER on AppTek Call-Center Dialogues across fourteen English accents, with the lowest error on eight of the fourteen. That is the most useful number in the launch for anyone running support lines, and it is the hardest to check.

• Noise and hallucination — 25.69% WER on HALAS, a human-annotated earnings-call set chosen for hallucination risk. The comparison set is vendors' own published figures: Qwen3-ASR 1.7B at 26.01%, Voxtral Small 24B at 27.35%, Nemotron 3.5 Transcribe at 31.59% and Granite Speech 5.0 2B at 34.51%. A 25.69% error rate is still one word in four, and the gap to the next model is 0.32 points.

• Style following — 5.05% preference-aware WER when the requested format is given as an instruction, second to IBM's Granite Speech 5.0 2B at 4.70%. The instruction-following claim is real but it is not a lead.

• Context biasing — on Contextual Earnings-22, keyword F1 rises from 71.9 with no hints to 96.8 given the clip's own keywords, and still scores 94.0 when handed the full hour's keyword list, most of which are distractors. This is the single most persuasive result in the launch, because distractor-heavy biasing is where naive keyword boosting usually falls apart.

The bill arrives at serving time

There is no hosted price for SOV-Transcribe-1. Salesforce published weights, not an endpoint, and no third-party inference host appears on the Artificial Analysis board with a per-1,000-minute figure for it. So the cost comparison is not dollars per minute against an API — it is GPU hours against an API, and the model card gives unusually complete numbers for that arithmetic.

• Throughput — with native vLLM support and the DFlash draft enabled, one RTX PRO 6000 Blackwell Server Edition (96 GB) transcribes up to 1,039 seconds of audio per second, and returns the transcript of a 30-second recording in 0.22 seconds. That is a batch-serving figure measured on a vLLM development build in November 2026, on one card.

• Latency by clip length — without the draft model: 0.16 s for up to 10 s of audio, 0.44 s for 10–30 s, 0.88 s for 30–60 s, 2.78 s for 1–5 minutes. With DFlash: 0.09, 0.15, 0.31 and 1.26 s.

• Memory — max_num_seqs=64 is required to keep startup profiling inside a 96 GB card once the draft model is loaded. vLLM's default of 1,024 concurrent sequences runs out of memory. This is a model that wants a dedicated card, not a slice of one.

• Comparison point — the cheapest proprietary entries on the Artificial Analysis board run at $1.67 to $5.00 per 1,000 minutes of audio. At 1,039 seconds of audio per second, a single card clears 1,000 minutes in about a minute of wall clock, so the break-even against a $5.00-per-1,000-minute API arrives quickly at volume — and never arrives at all if your volume is a few hours a week.

Three gaps the launch post does not dwell on

The first is streaming. SOV-Transcribe-1 is non-streaming by design. Google's Gemini 3.5 Transcribe has a separate live endpoint, OpenAI's GPT Transcribe accepts realtime transcription sessions over WebSocket, and Nemotron 3.5 Transcribe handles long recordings with timestamp output. If your product is live captions or a voice agent that needs partial transcripts, this model is not in the conversation, whatever its WER.

The second is speaker structure. The model card documents transcription, style control and context biasing. It does not document speaker diarization or word-level timestamps, and those are the two features that turn a transcript into something a compliance team or a meeting product can index. Gemini 3.5 Transcribe handles up to eight speakers; Nemotron 3.5 Transcribe advertises speaker-attributed transcription and timestamp prediction. SOV-Transcribe-1 asks you to recover that structure downstream.

The third is language coverage. Seven languages is narrow next to Nemotron's forty and Gemini's 85-plus, and it is narrow in a specific direction: all seven are European, and the multilingual benchmarks Salesforce published are consistent with that. If your audio is Portuguese and Dutch, this is arguably the best open model available. If it is Hindi, Japanese or Arabic, it is not an option.

Two smaller items belong on the same list. The technical report is linked as "coming soon", so the training data, the evaluation protocol beyond what the card shows, and the failure analysis are all still missing. And native vLLM support is a plugin today; the card says Salesforce aims to upstream it, which means the integration you adopt now may need to change.

When this is the right model anyway

The accuracy profile points at a specific buyer. Entity-heavy English audio — account numbers, email addresses, drug names, ticker symbols — is where the combination of 2.28% AA-WER and the Contextual Earnings-22 result actually pays. A pipeline that already knows the vocabulary of the call, and can hand it to the model as a hint list, gets a measurable lift that a general-purpose API will not match without the same biasing support.

Data residency is the second case. Apache 2.0 weights on your own hardware means the audio never leaves the building, and for call-centre recordings in regulated industries that is often the deciding constraint rather than a nice-to-have. The 30-minute limit and the absence of streaming are acceptable costs when the workload is overnight batch transcription of recorded calls.

Dictation with a house style is the third. Instruction control over verbatim versus cleaned-up output, digit formatting and filler removal is a genuine differentiator against Whisper-derived pipelines, where the prompt is a text prefix rather than an instruction the model was trained to follow.

One practical note for anyone wiring this into a larger system: the transcript is usually the input to something else — a summariser, a QA agent, a ticket classifier. That second hop is where a router earns its place. OrcaRouter puts 200+ text, image and video models behind one key with provider list pricing passed through at 0% markup, so the model that reads SOV-Transcribe-1's output can be swapped or failed over without a second contract. It does not host SOV-Transcribe-1 itself; the speech model runs on your GPU, and the routing starts where the text does.

What is still unproven

Three things will decide whether this model looks as good in six months as it does in its launch week. Independent reproduction of the entity and accent numbers is the first: VoiceCodeBench is private and AppTek is one vendor's corpus, so the claims that matter most commercially are the ones with the least outside evidence. Robustness on long, messy, multi-speaker audio is the second, and the 4.06% Earnings22 figure is the only public signal — real, but narrow. The third is whether the serving story gets simpler. A vLLM plugin pinned to one version, a 53 GB download and a 64-sequence cap are all friction that upstream support and quantised builds would remove.

For now the honest summary is that Salesforce has published the most accurate open speech recogniser available and has been unusually transparent about the shape of the trade: seven languages, thirty minutes, no streaming, one GPU. Whether that trade is right for you is a question about your audio, not about the leaderboard.