Cascading Layers of Competitive Associative Nets and Adaptive Vector Quantization Nets for Speaker Independent Word Recognition
A neural network based novel technique for speaker independent word recognition is introduced, which processes speech signal of spoken words by means of cascaded layers of competitive associative nets (CANs) and adaptive vector quantization nets (AVQNs). The CANs are employed for learning and recognizing linear prediction coefficients (LPCs) of speech signal, where the CANs provide stable learning and quick recognition of LPCs, namely the frame length for the present net to obtain LPCs has been set 5ms while that for the conventional algorithms usually requires 15ms to 30ms. In between the layers of CANs and AVQNs. layers of WTA cells and leaky integrators are placed, where the layer of WTA cells select a CAN which has the best associative matrix representing the LPC of speech signal in each frame, and the output of this layer is summed up by the leaky integrators for processing a word whose length of time is much longer than the signal processed by the CANs. Finally, the layer of AVQNs learns and recognizes the output of the leaky integrators at the end of the word signal. The adaptation scheme of the AVQN is useful in adapting to changing statistics of speakers, sensors, environments, etc. for speaker independent online word recognition. After showing preliminary experiments of vowel recognition, we illustrate the experiments of online spoken word recognition.
Shuichi Kurogi Masakazu Itakura Takeshi Nishida
Department of Control Engineering,Kyushu Institute of Technology, Kitakyushu 804-8550, Japan
国际会议
8th International Conference on Neural Information Processing(ICONIP 2001)(第八届国际神经信息处理大会)
上海
英文
431-436
2001-11-14（万方平台首次上网日期，不代表论文的发表时间）