Fixed-point arithmetic trade-offs in adaptive filters for speech recognition

In this paper the data precision behavior of adaptive filters is studied when implemented in fixed point. The performance evaluation has been tested under two parameters: the number of bits used for the integer and fractional parts in fixed point format and two's complement arithmetic and the preprocessing order of speech data. To validate our results we have evaluated the relative mean square error between the results obtained using the standard 32-bit floating point and different fixed point formats. The input data being used for testing are speech sounds in Spanish for speech recognition purposes. The obtained results allow us to optimally dimension an arithmetic unit in fixed-point formats that can be quickly implemented by high-level synthesis tools. The results from this work can also be used for implementing the pre-processing stage for begin-end word detection.