What changed
Salesforce released SOV-Transcribe-1 under Apache 2.0, a non-streaming automatic-speech-recognition model with 4 billion active and 26 billion total parameters. The weights support English and multilingual transcription, context biasing and output-style steering.
Why it matters
Teams can now run and adapt a large speech model under a permissive license instead of sending audio to a hosted API. Local control is useful for sensitive audio, but deployability depends on actual hardware, language and noise performance.
What to do
Reproduce word and named-entity error rates on consented private audio, measure memory and real-time factor on target hardware, and verify the exact language, accent and punctuation behavior before replacing a production recognizer.
Caveat
The release is non-streaming, and the published benchmark comparisons are compiled or reported by Salesforce. A technical report was still forthcoming, so architecture, data and subgroup limitations are not fully documented.