Salesforce releases open-weight transcription AI, claiming a benchmark lead

Salesforce released SOV-Transcribe-1 on November 12, 2026, making its speech-recognition weights available on Hugging Face under Apache 2.0. The company says it leads open-weight models on Artificial Analysis’ non-streaming transcription ranking. Beyond general accuracy, the release targets a practical problem: getting names, email addresses and numbers right enough for software to use.

The release gives developers downloadable model weights for building dictation tools, voice agents and call transcription on infrastructure they control. It is built for non-streaming speech recognition, not live streaming transcription. Salesforce describes the model as having 4 billion active parameters and 26 billion total; those figures distinguish the portion used during processing from the model’s full size.

A lead among downloadable models, not every system

Salesforce places SOV-Transcribe-1 first among open-weight systems and seventh overall on Artificial Analysis, with rankings dated November 12. Its reported word error rate is lower than those of NVIDIA’s Nemotron 3.5 Transcribe and Mistral’s Voxtral Small. The comparison is also a model-size pitch: Salesforce lists Nemotron at the same active parameter count and Voxtral at 24 billion active parameters, six times SOV’s stated figure.

2.28%SOV-Transcribe-1

Salesforce’s reported result for its new model. Lower word error rates are better.

2.81%Nemotron 3.5 Transcribe

The NVIDIA comparison score cited in Salesforce’s release.

2.77%Voxtral Small

The Mistral comparison score cited in Salesforce’s release.

A separate evaluation asks a stricter question than whether a transcript is mostly readable. VoiceCodeBench tests details such as phone numbers, email addresses, URLs and code symbols. Its task-success measure passes a recording only when every target detail is correct. Salesforce reports that SOV passes 76.3% of recordings while recovering 93.4% of individual target details, placing it first among open-weight models and second among 23 systems overall on both measures.

That distinction matters for the applications Salesforce is targeting. A dropped underscore or swapped digit can leave a transcript looking plausible while producing an incorrect software argument or an invalid database field. The two VoiceCodeBench scores separate how often individual details survive from how often a recording gets all of them right.

Instructions change how speech becomes text

The model’s controls cover both presentation and recognition. Salesforce’s release describes two ways developers can steer it without changing the underlying weights:

• Style instructions: The default is verbatim transcription. Natural-language requests can instead remove fillers or write numbers as digits, adapting the transcript to a dictated message rather than a word-for-word record.
• Vocabulary context: Developers can supply keywords or a short description of the audio to help resolve rare or similar-sounding terms, such as a customer’s name or a drug name.

Salesforce tested vocabulary hints on Contextual Earnings-22, which uses 15-second earnings-call clips centered on names of people, companies or products. Supplying only the keywords spoken in a clip raised keyword F1—a score for recognizing those names accurately—from 71.9 to 96.8. Giving the model the entire call’s keyword list, including many distractors, produced 94.0, the highest open-weight result in Salesforce’s comparison.

Harder audio produces very different error rates

The launch also includes tests beyond the headline ranking. On AppTek Call-Center Dialogues, covering 14 English accents, Salesforce reports an average word error rate of 8.21% and the lowest rate on eight accents. On HALAS, an earnings-call benchmark focused on audio likely to trigger invented words, its reported error rate is 25.69%. That beats the other open-weight models listed, but remains a substantially harder test than the headline result.

Nor does SOV lead every test Salesforce presents. On Preference-ASR, which checks whether transcripts follow requested formatting, its instruction-guided error rate is 5.05%, behind IBM Granite Speech 5.0 2B’s 4.70%. Salesforce marked its technical report as coming soon when it published the release.