Ernst Bunge

Automatic speaker recognition by computers

Summary form only given, as follows. This paper describes a new modular speaker recognition system consisting of a st of real?time speech analysis processors and a pattern recognition software package. Within a government sponsored research project, combinations of different speech analysis procedures and different pattern recognition algorithms are compared in order to find optimal subsystems, to be applied to security systems or law enforcement, for given boundary conditions. In order to find the influence of different techniques, distance measures, quantisation band distortions on the recognition rate of given data base (2,500 utterances), a study has been carried out, results of which are being discussed.

Method and device for verifying signals, especially speech signals

The characteristic values are obtained from the signal to be verified, e.g. in the case of a voice signal the total energy in the individual frequency spectrum ranges, and these characteristic values are quantified. The frequency with which the individual partial ranges of the quantified characteristic values occur is found during a sampling phase, and these frequencies are stored. In the test phase, the signal is analyzed in the same way and the stored frequencies of the partial ranges into which the characteristic values of the signal to be verified have fallen are multiplied and compared with a threshold. In an arrangement for implementing the method, there is a characteristic store with a store address for a multi-digit binary number for each partial range of each characteristic. This store is addressed by a counter, the positions of which are allocated to the individual characteristics or spectrum ranges of the voice signal, and by the numbers of the partial ranges of the characteristics found during the analysis into which the characteristic values of the signal to be analyzed have fallen. In the sampling phase the content of each addressed store address is raised by 1 and in the test phase the frequencies obtained in the store addresses addressed are multiplied together. When a voice signal is processed for verification of the speaker, with the stepwise formation of the long-term spectrum, the intermediate values are compared with a range limit of a partial range read out from a range limit store and increased by 1 on exceeding a number allocated to this characteristic in an intermediate store. In this manner, it is possible at the end of the voice signal, immediately and without any further processing steps to obtain the quantified characteristics in the form of the numbers of the partial ranges which can then be used to increase the frequencies in the characteristic store. A summation over all spectrum values is made to standardize the incoming voice signals. Subsequently, to all spectrum values is added a given fraction of their value, performed in the simplest manner by position shifting, until the sum of all spectrum values exceeds a predetermined constant. Thus the determination of the quantified characteristics or the partial range sums is immediately effective. The multiplication of the frequencies may be replaced by the addition of the logarithmic frequencies.

Method and arrangement for sound analysis

For sound analysis, especially for speaker identification, a long-term spectrum is often formed from the recorded signal and compared to a long-term spectrum of a previously recorded signal. When the signal is, for example, received over telephone lines the transfer function thereof may distort the received signal to such an extent that identification is impossible. It is now proposed not to use the long-term spectrum for identification but rather its change, that is to say to accumulate the difference between two consecutive short-term spectra or the difference between such short-term spectrum and the long-term spectrum and to divide this accumulated value by the long-term spectrum in a Divider. Since the numerator and denominator each contain the frequency-weighting factor (aJ) of the telephone line transfer function, this factor is effectively cancelled out by virtue of the division operation performed in the Divider thereby eliminating the effect of the unknown transfer function. As a result the influence of the transfer function of the transmission path disappears almost completely.

Statistical techniques for automatic speaker recognition

Within a government sponsored research program various methods of speech analysis techniques and pattern recognition methods have been applied to the speaker identification and verification problem. For this purpose a modular speaker recognition system has been developed to be used for comparative studies. Real-time speech signal analysis, mainly based on two-stage statistical measurements in combination with minimum risk classifiers allows code-word related as well as text-independent speaker verification and identification, both with very high accuracy for male and female voices. This paper describes the structure and modules of the speaker recognition system, results of comparative experiments are being discussed.