On-Device Speech Recognition

A smaller model for local hardware

The rebuild brought a recent cloud-model generation onto consumer laptops. Speechmatics describes quantization and inference optimization as central to meeting Adobe’s resource limits. This was its own model development, not a renamed Whisper package.Speechmatics: Rebuilding the on-device speech engine

The company’s later access guide distinguishes an embedded library from an on-premises server. With the local library, audio and transcripts can remain on the user’s device. The application developer distributes model updates with the application.Speechmatics: On-device early access and integration

Practical guidance. Local recognition is useful for editing without a network connection. Its privacy benefit depends on what the rest of the application does with the resulting text.Speechmatics: On-device early access and integration

Revision dates and public naming

The public account names a new model in April 2026 but does not give it a stable checkpoint identifier. This article covers the documented on-device family; it does not turn each hardware optimization into another model.Speechmatics: New on-device model in Adobe Premiere