Speech interface VLSI for car applications

A user-friendly speech interface for car applications is highly needed for safety reasons. This paper will describe a speech interface VLSI designed for car environments, with speech recognition and speech compression/decompression functions. The chip has a heterogeneous architecture composed of ADC/DAC, DSP, RISC, hard-wired logic and peripheral circuits. The DSP not only executes acoustic analysis and output probability calculation of HMMs for speech recognition, but also does speech compression/decompression. On the other hand, the RISC works as a CPU of the whole chip and Viterbi decoder with an aid of hard-wired logic. An algorithm to recognize a mixed vocabulary of speaker-independent fixed words and speaker-dependent user-defined words in a seamless way is proposed. It is based on acoustic event HMMs which enable a template creation from one sample utterance. The proposed algorithm embedded in the chip is evaluated. Promising results of the algorithm for multiple languages are shown.