Though the Gaussian estimator was designed for Sphinx 3 and the MIPS-like embedded processor, the results are widely applicable to other architectures and recognizers. There are several levels at which this system may be integrated into a speech recognition task pipeline similar to Phased. For example, an intelligent microphone may be created by using a simple low power DSP to handle the A/D conversion and FE phase, and then a GAU coprocessor attached to the DSP may be used for probability estimation. The probability estimates can then be sent to a high-end processor or custom accelerator that does language model computation. The GAU coprocessor can then hide more than 50% of the compute effort required for speech recognition. On desktop systems, the Gaussian accelerator may be part of a sound card or the Gaussian accelerator may be directly attached to the main processor. On commercial voice servers, the Gaussian estimator may be directly built into the line cards that interface to the telephone network thereby freeing up server resources for language model and application processing. This also has important implications for server scalability, discussed in the Section 6.5.2.