On November 12, 2026, Salesforce released SOV-Transcribe-1, its first open-weight speech recognition model, with the weights published the same day on Hugging Face under the Apache 2.0 license. On Artificial Analysis' non-streaming speech-to-text leaderboard, it posts a 2.28% word error rate, ranking first among all open models and seventh overall, ahead of Mistral's Voxtral Small and NVIDIA's Nemotron 3.5 Transcribe.
What 2.28% Means: First Among Open Models, Seventh Overall
Word error rate is the standard hard metric for speech-to-text — lower is better. SOV-Transcribe-1's 2.28% beats Voxtral Small's 2.77% and Nemotron 3.5 Transcribe's 2.81%, placing seventh among 44 evaluated models, with only a few closed models ahead. On VoiceCodeBench it ranks second among 23 systems: 76.3% of recordings have every key entity transcribed correctly, and entity recall is 93.4%. In customer-service dialogue tests across 14 English accents, its average word error rate is 8.21%, the lowest among the evaluated models for 8 of those accents. The model has 26.6 billion total parameters and activates only 4.6 billion at a time.
More Than Accuracy: Format Control on Instruction, Rare Words Recognized
Beyond accuracy, SOV-Transcribe-1 targets the real pain points of enterprise recordings. It supports instruction-controlled transcription formats: a court record must keep every filler word verbatim, while a dictated message should read cleanly — the same recording can be rendered differently on request. Context biasing lets users supply keywords or a short audio brief in advance, improving transcription of hard names and drug names; giving the clip's keywords lifts keyword F1 from 71.9 to 96.8. On real earnings-call audio that tends to trigger hallucinations, its word error rate is 25.69%, lower than all open-weight comparisons.
The model belongs to Salesforce's sovereign AI program, and its target customers are clear: government agencies and regulated enterprises that cannot send recordings to someone else's servers. It supports local deployment, with weights downloadable to run on an organization's own infrastructure, and its main use cases are meeting notes, customer-service calls, and podcast recordings. But the limits are equally clear: it does non-streaming transcription only, so it does not fit live captioning; it supports just seven European languages — English, French, German, Spanish, Italian, Portuguese, and Dutch — with no Chinese; the weights run to about 53 GB, so deployment needs matching GPU memory and inference setup, and the technical report has not yet been published.
For Chinese users, SOV-Transcribe-1 is more of a reference point: the open speech-recognition bar has been pushed down to 2.28%, while Chinese-language scenarios still await a comparable model that supports Chinese. What to watch next is independent re-testing once the technical report is published, and the model's long-term stability on real customer-service recordings.
