Robust speech recognition for mobile applications

This paper proposes a robust speaker-independent, connected digit recognition system for mobile applications. The system requires a small amount of ROM and low computational cost with a high recognition accuracy. In addition, the system can be efficiently implemented on most currently available 32-bit fixed-point DSP chips. To reach these goals, we combined robust speech parameter processing technologies with dual matrix quantization (MQ) and vector quantization (VQ) pairs, which supply discrete gender-dependent HMM to increase the performance of HMMs. The dual MQ/VQ pairs exploit the “evolution” of the speech short-term spectral envelopes with one pair providing error compensation using LSP mean compensated coefficients. In a car noise environment, the system attains an 80% average connected digit recognition accuracy around 10 dB. A digit accuracy of 93% is obtained at 5 dB