| In recent years, the recognition technology of audio signal, including speech recognition, mechanical noise recognition, underwater acoustic signal recognition, have been achieving great progress.In the field of audio recognition, with many mature and creative technologies applying, especially the Hidden Markov Models (HMM), the effect and efficient of the audio recognition system have been enhanced. But due to the mismatch between training and testing environment (such as background, audio transition channel), the recognition systems based on HMM tends to drastically degrade in performance. For the reason of the theory of human aural perception, in this paper, we present a system based on the fusion of multi-band Continuous HMM (CHMM) and Back Propagation Neural Network (BPNN). The multi-band CHMM system is composed of several sub-bands and a full-band. We derive appropriate feature for each band and train independent recognizers for each subbband region. Then, we use BPNN to make the final decision by combined all CHMM system's outputs and frame average energy. Through the judgment and syncretism ability of BPNN, the system can effectively handle audio signals in different conditions. The simulations are completed in Matlab platform and the results in simulations are compared with other relevant algorithms.In addition, we propose an embedded audio recognition system based on FPGA and Nios II. This scheme uses Cyclone FPGA chip and combines software and hardware design. FPGA completes several tasks such as pre-emphasis, windows and frame. Nios II soft core fulfills endpoint detection, feature extraction, discipline, recognition, input control and output display, etc. The audio signal feature, in this scheme, is the LPC Mel Cepstrum Coefficient (LPCMCC) and recognition algorithm is Dynamic Time Warping (DTW). In the debug process, audio data is gotten through RS232 from PC. This paper completes the development and debug of system, and achieves the requirements of design. |
PDF Full Text Request