An improved in-wall speaker assembly providing better performance and less intrusion than prior art speakers. The speaker assembly is oriented such that the driver fires vertically, parallel to the wall studs, rather than horizontally and perpendicular to the wall studs. The cone and surround of the speaker assembly are generally rectangular and/or conform to the area available between studs. Additionally, the cone is not attached to the voice coil, but is connected to the voice coil by a shaft which transmits the energy from the voice coil to the cone. In another aspect of the invention, the cabinet or housing includes two housings.