Superwhisper trains their own S1 models for this, you can use either a smaller local version of them or a beefier one in the cloud