What happened
Salesforce’s Sovereign Intelligence team introduced SOV-Transcribe-1 on November 12, 2026. The announcement describes it as the team’s first open-weight speech-augmented language model, designed for non-streaming automatic speech recognition. The model has 4 billion active parameters and 26 billion total parameters. Salesforce says the weights are available on Hugging Face under the Apache 2.0 license, for developers and enterprises building voice agents, dictation products, and call-transcription systems. The announcement does not describe a hosted service or a price; its stated release is the downloadable weights.
The model is intended to do more than transcribe fluent speech. Salesforce highlights entity handling for details such as email addresses, phone numbers, account numbers, URLs, and code symbols, where a small transcription error can change a downstream action. On VoiceCodeBench, a benchmark Salesforce identifies as BeSimple’s private leaderboard, SOV-Transcribe-1 passes 76.3% of recordings under a measure that requires every target entity to be correct, and recovers 93.4% of entities. Salesforce says those results place it first among open-weight models and second among 23 systems overall on both measures.
For general transcription, Salesforce reports a 2.28% word error rate on Artificial Analysis and says the model ranks first among open-weight systems for non-streaming speech-to-text and seventh overall. In that comparison, it reports 2.81% for NVIDIA’s Nemotron 3.5 Transcribe with the same number of active parameters, and 2.77% for Mistral’s Voxtral Small, which it says has six times as many active parameters. The article says similarly sized open-weight models are around 4% WER. On AppTek Call-Center Dialogues, a long-form benchmark spanning 14 English accents, Salesforce reports an average 8.21% WER and the lowest WER on eight accents: Australian, British, Scottish, Welsh, Irish, Indian, Mexican, and Singaporean English.
The announcement also describes instruction-controlled formatting. SOV-Transcribe-1 transcribes verbatim by default, but can follow natural-language instructions to, for example, write numbers as digits or remove disfluencies. On Preference-ASR, Salesforce reports a 5.05% preference-aware WER with instructions, behind IBM’s Granite Speech 5.0 2B at 4.70% and ahead of Qwen3-ASR 1.7B at 5.96% and Nemotron 3.5 Transcribe at 6.16%.
For difficult or domain-specific terms, developers can provide keywords or a short audio description to bias transcription. In Salesforce’s Contextual Earnings-22 test, clips are transcribed with no keywords, clip-specific keywords, or the full set of keywords from an hour-long call. With clip-specific keywords, the model’s keyword F1 rises from 71.9 to 96.8; with the full keyword list, including distractors, it scores 94.0. Salesforce also reports a 25.69% WER on HALAS, a human-annotated benchmark of real earnings-call audio, lower than the other open-weight models listed in the announcement.
What the archive adds
New Runtime covered NVIDIA’s Nemotron 3.5 Transcribe, an open-weight speech-recognition model, on October 31, 2026. SOV-Transcribe-1 is a new release in the same open-weight ASR category, with Salesforce presenting its own comparative benchmark results and a focus on entity accuracy and controllable formatting.
Trend signal
This does not yet fit a pattern New Runtime tracks. Salesforce’s announcement concerns a speech-recognition model and self-hostable weights; it does not establish a broader shift in generative-media infrastructure or demonstrate the replacement of existing voice stacks.
Why now
Salesforce published the announcement on November 12, 2026, and says the model weights are available today. Its release positions control of infrastructure as a reason to build voice experiences with open weights, while the reported benchmarks provide a basis for teams to evaluate the model against other open systems.
Why it matters
Transcription errors can propagate into actions: a wrong email address, digit, or product name can invalidate a database field or change a voice agent’s tool call. If the reported entity scores hold on a team’s own audio, an open-weight model that supports context biasing could make it easier to keep this step inside a controlled deployment and adapt it to a specialized vocabulary.
The trade-off is that model weights do not by themselves provide a production voice pipeline. Teams still need to run inference, manage audio capture and storage, monitor errors, and determine how transcription output is checked before an agent acts on it. Since SOV-Transcribe-1 is described as non-streaming, it should be evaluated for batch and post-call workflows rather than assumed to support live interaction. Its value will depend on measured performance for the accents, recording conditions, and entities that matter in each deployment.
Limits and uncertainty
The announcement’s rankings and error rates are claims made by Salesforce, and the article does not provide an independent evaluation. Its results are benchmark-specific; they do not establish performance on every language, recording environment, or enterprise vocabulary. The VoiceCodeBench leaderboard is described as private, so readers cannot inspect its full methodology from the announcement. The release also does not establish inference cost, hardware requirements, throughput, or a supported streaming mode. A technical report is described as coming soon, but the announcement does not include it.
Skill to build now
Take a representative sample of your own calls or dictated audio and make a small evaluation set that labels names, account identifiers, email addresses, and other fields where a single character matters. Compare SOV-Transcribe-1 with your current recognizer using both ordinary transcription and context keywords. Measure exact entity recovery as well as word error rate; do not let a strong average score obscure failures on the fields that trigger actions.
Reader takeaway
SOV-Transcribe-1 gives teams another Apache 2.0 option for self-hosted, non-streaming speech recognition. Salesforce’s strongest claims concern recovery of critical entities and keyword biasing, but the numbers are vendor-reported. Test it on your actual audio and measure whether it gets the fields your workflow depends on right.