Tomofumi Nakatani

Binary processing of sound spectrum

A frequency spectral distribution of sound signal has a plurality of discrete data at respective channels where sampling takes place. Such a spectral distribution, defining a parallel signal, is converted into a time series signal which is used to produce a peak-detected signal and a zero-cross signal. Then, based on the timing of the zero-cross signal, the peak-detected signal is compared with the original parallel signal to process the original parallel signal to be binary-valued. The present invention provides a simple hardware structure having the above-described function.

Voice recognition system and method involving registered voice patterns formed from superposition of a plurality of other voice patterns

After sampling a voice signal by a microphone (31), a voice interval is detected by a voice interval detector (32). Then, the voice signal is processed by a frequency analyzer (33) having a predetermined number of channels at a predetermined time interval, whereby a corresponding portion of the voice signal is quantized at each channel. Then, quantized data obtained from these channels are binary converted (34) to thereby form a frame comprised of a series of binary data. Each data of this frame corresponds to one of the channels, and preferably it is set to be an integral multiple of a computer calculation unit (e.g., 4 bits, 8 bits, etc.). When forming a combination frame by superimposing two or more of such frames, the combination frame is divided into layers, whereby it is so structured that each bit is represented by a binary number in each layer. On the other hand, it may also be so structured that one frame is divided in a plurality of sub-frames and a preliminary comparison process is carried out using each sub-frame.

Voice spectrum analyzing system and method

A voice spectrum analyzing method and system subjects a voice signal to frequency analysis by passing the voice signal through a filter bank containing a plurality of band pass filters each having a different band pass filter range to produce a voice spectral pattern over a predetermined frequency range. The voice spectral pattern is sampled at a predetermined sampling time interval successively, and each of the sampled voice spectral patterns is stored for a predetermined time period. While the voice spectral pattern is retained, it is scanned once in a predetermined time sequence, thereby forming an approximate periodic signal from a series of the scanned voice spectral voice patterns. And the periodic signal thus formed is then digitized and filtered through a digital high pass filter, thereby obtaining a high frequency component from the periodic signal which indicates a feature parameter of the voice signal.

Speaker-independent word recognition using fuzzy pattern matching

Abstract A new pattern matching algorithm, employing fuzzy logic for speech recognition, is proposed. In the field of speech recognition, there are two problems to overcome before realizing a speaker-independent speech recognition machine. They are firstly, fluctuation of pattern length, and secondly, that of frequency, caused by the variance of utterance speed, and differences of resonance frequency of the vocal tract, respectively. In the described algorithm these fluctuations are translated into membership functions. The membership functions are regarded as templates for pattern matching. In this paper, after the basic theory of recognition is described, the theory is modified to a form easily used by machine. As an experimental result a recognition rate of 93.2%, enrolling 120 words, was achieved in the case of Japanese. In the case of other languages 92.8% and 93.7% were achieved for English and German, respectively. These results indicate this method is realistically accurate. It seems to be worth noting that these results were achieved using a simple algorithm and accordingly the machine is small and inexpensive.