Salesforce releases SOV-Transcribe-1, an open-weight speech-to-text model

Salesforce has released SOV-Transcribe-1, an open-weight speech-to-text model under the Apache 2.0 licence, its sovereign AI team announced on Hugging Face. Salesforce says it ranks first among open-weight models for non-streaming speech-to-text on Artificial Analysis and seventh overall, as of 12 November.

The model has 4 billion active parameters. Salesforce reports a 2.28% word error rate on Artificial Analysis's benchmark, against 2.81% for NVIDIA's Nemotron 3.5 Transcribe and 2.77% for Mistral's much larger Voxtral Small. On a call-centre benchmark covering 14 English accents it reports an 8.21% average error rate, the lowest on 8 of the accents, and on a benchmark of real earnings-call audio it says it beats other open-weight models, including Qwen3-ASR and IBM's Granite Speech.

SOV-Transcribe-1 works in English, French, German, Spanish, Italian, Portuguese and Dutch. It can follow instructions on formatting, such as keeping filler words or writing numbers as digits, and accepts keywords or a short description to help it get rare names and terms right. Salesforce pitches it for voice agents, dictation and call transcription that companies run on their own infrastructure. The benchmark figures are Salesforce's own.