Salesforce released SOV-Transcribe-1, a new open-weight speech-to-text model. It ranks as the top-performing model in its class. Developers and enterprises can download the model from Hugging Face. It uses an Apache 2.0 license, allowing deployment on their own infrastructure.
Independent benchmarks from Artificial Analysis show SOV-Transcribe-1 achieved a 2.28% word error rate. It ranked first among 16 open-weight models. It also placed seventh overall out of 64 systems evaluated, including proprietary ones. The model accurately captures details such as names and numbers. It handles non-streaming transcription tasks.
SOV-Transcribe-1 is a speech-augmented language model. It features 26.6 billion total parameters, with 4.6 billion active during use. This release supports Salesforce's Sovereign AI initiative. The initiative serves government and enterprise clients requiring high data security and control.