Salesforce's SOV-Transcribe-1 leads open-weight batch speech-to-text. It does not stream, and messy audio still breaks often

On 12 November 2026, Salesforce's Sovereign Intelligence team released SOV-Transcribe-1, an Apache 2.0 speech model, with a blog post and a model card on Hugging Face. The headline is that it ranks first among open-weight models for non-streaming speech-to-text on Artificial Analysis, and seventh overall, as of that date, with a 2.28% score on Artificial Analysis's word-error index. Non-streaming, also called batch transcription, means the model is built to transcribe a finished recording. It is not built to emit words while someone is still speaking.

That distinction matters more than the ranking. The blog advertises "real-world speech" and a 4 billion active-parameter size that sits on a favorable error-versus-size chart. The model card is more precise: 26.6 billion parameters in total, 4.6 billion of them active, including the audio encoder. In a Mixture-of-Experts model, only a subset of the network fires for each piece of input, but you still have to load the full 26.6 billion parameters to run it. Salesforce's own serving notes assume a 96 GB class GPU for the vLLM setup it recommends with its draft model.

The architecture is a speech-augmented language model, not a standalone speech encoder. The model card describes a Whisper large-v3 audio encoder, an audio projector, and Google's Gemma 4 26B A4B Mixture-of-Experts language model. Audio is expected at 16 kHz. Optional text instructions and a short context can be supplied with the recording. The card lists English, French, German, Spanish, Italian, Portuguese, and Dutch. The longest audio Salesforce says it has tested is 30 minutes. The default vLLM transcription endpoint, unless a setting is raised, accepts 600 seconds per request, which is 10 minutes, not 30.

Native support in vLLM and Hugging Face Transformers is still being upstreamed. Until that lands, the repository ships a vLLM 0.32.0 plugin and modeling code for Transformers. Salesforce recommends greedy decoding. It also warns that the model can, in rare cases, repeat a character or phrase until it hits the token limit, and that it removes those loops with a post-processing step adapted from the Qwen3-ASR toolkit.

The serving numbers are vendor measurements on an NVIDIA RTX PRO 6000 Blackwell GPU with 96 GB of memory. With the DFlash draft model, which proposes 15 tokens at a time for the main model to check, Salesforce reports that a recording of one to five minutes is transcribed in 1.26 seconds, one request at a time, and that shorter clips finish in a fraction of a second. A later note, measured on a November 2026 vLLM development build with native support, says one such GPU transcribes up to 1,039 seconds of audio per second in BF16 with the draft, and returns a 30-second transcript in 0.22 seconds. Those are offline throughput figures, not a live-captioning latency story. The card also says that vLLM's default of 1,024 concurrent sequences runs out of memory with the draft, and that setting the cap to 64 keeps start-up memory profiling within a 96 GB GPU.

On accuracy, the blog places SOV-Transcribe-1 at 2.28% on Artificial Analysis's word-error index, ahead of NVIDIA's Nemotron 3.5 Transcribe at 2.81% with a similar active-parameter count, and ahead of Mistral's Voxtral Small at 2.77% with far more active parameters, 24 billion in the blog's comparison. Other open-weight models of a similar size, the blog says, sit around 4%. I have not independently re-run Artificial Analysis's index. The ranking should be read as Salesforce reporting a third-party leaderboard, as of 12 November 2026.

The more distinctive claims are about entities, instructions, and context, not average word error on clean speech.

VoiceCodeBench, which the blog identifies as BeSimple's private leaderboard, scores a recording as a success only if every target entity in it is correct, and also measures how many of those entities are recovered. Salesforce reports a 76.3% task success rate and 93.4% entity recovery, first among open-weight models and second among 23 systems overall on both measures. Because the leaderboard is private, those figures cannot be checked the way a public Artificial Analysis page can.

On AppTek Call-Center Dialogues, a long-form set of agent and customer calls spanning 14 English accents, the blog reports an 8.21% average word-error rate, and the lowest error on eight of those 14 accents.

On HALAS, a human-annotated set of real earnings-call audio chosen because models are likely to invent words there, SOV-Transcribe-1 is reported at 25.69% word error, only 0.32 points ahead of Alibaba's Qwen3-ASR 1.7B at 26.01%, and further ahead of Voxtral Small 24B, Nemotron 3.5 Transcribe, and IBM's Granite Speech 5.0 2B. A 25.69% word-error rate still means roughly one word in four is wrong on this set. If "real-world" means a noisy earnings call, that number does not support a strong robustness claim.

Style control is instruction-following. The model transcribes verbatim by default and can be told, in ordinary language, to write numbers as digits or to drop fillers. On Preference-ASR, which scores whether a transcript follows a requested format, Salesforce reports 5.05% preference-aware word error with instructions, second to Granite Speech 5.0 2B at 4.70%. On Contextual Earnings-22, giving the model the keywords actually spoken in a 15-second clip raises keyword F1 from 71.9 to 96.8. Giving it every keyword from the full hour-long call, most of them distractors, still scores 94.0, the highest among the open-weight models in the blog's table.

The gaps that should travel with the "#1 open-weight" line are therefore specific.

The model does not stream. A voice agent that must show words as the caller speaks still needs a different system, or a still-unreleased streaming variant. The 4.6 billion active-parameter chart is real as a compute-per-token story only if you already have the memory and the disk footprint of a 26.6 billion parameter Mixture-of-Experts model, built on Gemma 4 and Whisper large-v3. Salesforce's own vLLM notes point to a 96 GB class card for the recommended draft setup. That is a data-center or high-end workstation cost, not a laptop one.

Several of the most advertised scores, including VoiceCodeBench, AppTek, HALAS, Preference-ASR, and Contextual Earnings-22, are reported by Salesforce, not reproduced here from an independent page. VoiceCodeBench is explicitly private. HALAS, even on Salesforce's numbers, shows that messy financial audio remains hard for every open model in the comparison, with only a 0.32 point gap over Qwen3-ASR 1.7B. The model card also documents a failure mode, repetition loops, that needs extra cleanup after decoding.

The fair counter is that average word error on clean speech was already a weak way to judge enterprise transcription, and Salesforce is measuring things that actually break downstream software: an email address, a phone number, a drug name, a requested number format. If the Artificial Analysis rank holds, SOV-Transcribe-1 is the open-weight model to beat for offline English-heavy batch work, under an Apache 2.0 license, which means a company can run it on machines it controls. For hospitals, banks, and public agencies that cannot send call audio to a US API, that is the product. Instruction-controlled formatting and keyword hints are also a practical answer to the usual mess of "the transcript is mostly right, but the account number is wrong."

It is also fair to say that a 26.6 billion parameter model that finishes a five-minute recording in about a second, on the hardware Salesforce used, is fast enough for many offline queues. The limitation is the job it was built for, not a hidden failure to meet a speed target it never claimed.

Over the next six months, the test is whether Artificial Analysis still shows SOV-Transcribe-1 ahead of Nemotron 3.5 Transcribe and Voxtral Small on the public non-streaming index, and whether anyone publishes an independent VoiceCodeBench-style entity test that is not a private leaderboard. Watch also whether Salesforce, or a rival, ships an open-weight streaming model with similar entity and instruction control, and whether teams outside Salesforce can run the 26.6 billion parameter stack at the quoted speed on the same class of GPU without the still-to-be-upstreamed vLLM path. If HALAS-like audio stays near 26% word error, the honest use case is controlled dictation and call recordings with a human in the loop, not hands-off transcripts of noisy financial speech.