Speaker Recognition Experiments Research Articles

Learning statistically efficient features for speaker recognition

We apply independent component analysis for extracting an optimal basis to the problem of finding efficient features for representing speech signals of a given speaker. The speech segments are assumed to be generated by a linear combination of the basis functions, thus the distribution of speech segments of a speaker is modeled by adapting the basis functions so that each source component is statistically independent. The learned basis functions are oriented and localized in both space and frequency, bearing a resemblance to Gabor wavelets. These features are speaker-dependent characteristics and to assess their efficiency we performed speaker recognition experiments and compared our results with the conventional Fourier basis. Our results show that the proposed method is more efficient than the conventional Fourier-based features, in that they can obtain a higher recognition rate and coding efficiency.

The correlation between auditory speech sensitivity and speaker recognition ability

In various applications of forensic phonetics the question arises as to how far aural-perceptual speaker recognition performance is reliable. Therefore, it is necessary to examine the relationship between speaker recognition results and human perception/production abilities like musicality or speech sensitivity. In this study, performance in a speaker recognition experiment and a speech sensitivity test are correlated. The results show a moderately significant positive correlation between the two tasks. Generally, performance in the speaker recognition task was better than in the speech sensitivity test. Professionals in speech and singing yielded a more homogeneous correlation than non-experts. Training in speech as well as choir-singing seems to have a positive effect on performance in speaker recognition. It may be concluded, firstly, that in cases where the reliability of voice line-up results or the credibility of a testimony have to be considered, the speech sensitivity test could be a useful indicator. Secondly, the speech sensitivity test might be integrated into the canon of possible procedures for the accreditation of forensic phoneticians. Both tests may also be used in combination.

An Interactive Technique for Matching Speaker Identity

Abstract A speaker recognition experiment is described where subjects are asked to identify 1 out of 12 unknown speakers from short samples of speech. All were male speakers without strong regional accents and unknown to the subject group. The task was to match the voice of a test utterance to one of the speakers. The test was run individually as an interactive computer-controlled task. Results on speed and precision of this task are presented. The result demonstrates that it is possible for subjects to quite accurately match unknown voices, speaking different utterances, if the test utterances are reasonably long. However, listeners differ widely in this ability. The study will be used as a baseline for synthesis of speaker characteristics.

Evaluation comparative du pouvoir discriminant d'indices de (dis)similarité spectrale

In this paper, the discriminatory abilities of 16 similarity and dissimilarity indices are tested by means of a speaker-recognition experiment involving high-dimensional longterm speech spectra. The results give the ranking of the 16 indices on the basis of their discriminatory abilities

Speaker recognition using neural networks and conventional classifiers

An evaluation of various classifiers for text-independent speaker recognition is presented. In addition, a new classifier is examined for this application. The new classifier is called the modified neural tree network (MNTN). The MNTN is a hierarchical classifier that combines the properties of decision trees and feedforward neural networks. The MNTN differs from the standard NTN in both the new learning rule used and the pruning criteria. The MNTN is evaluated for several speaker recognition experiments. These include closed- and open-set speaker identification and speaker verification. The database used is a subset of the TIMIT database consisting of 38 speakers from the same dialect region. The MNTN is compared with nearest neighbor classifiers, full-search, and tree-structured vector quantization (VQ) classifiers, multilayer perceptrons (MLPs), and decision trees. For closed-set speaker identification experiments, the full-search VQ classifier and MNTN demonstrate comparable performance. Both methods perform significantly better than the other classifiers for this task. The MNTN and full-search VQ classifiers are also compared for several speaker verification and open-set speaker-identification experiments. The MNTN is found to perform better than full-search VQ classifiers for both of these applications. In addition to matching or exceeding the performance of the VQ classifier for these applications, the MNTN also provides a logarithmic saving for retrieval.

SDDD: A new dissimilarity index for the comparison of speech spectra

SDDD, an index for the comparison of speech spectra, is introduced in this paper. Its discriminatory ability is established by comparing SDDD with the best classical index on the basis of a speaker-recognition experiment using Long Term Average Spectra.

Text‐independent speaker recognition experiments using codebooks in vector quantization

A text‐independent speaker clustering approach to speaker‐indepencent speaker recognition through vector quantization (VQ) was investigated, where the distortion value was used as a clustering measure. To show the possibility of the text‐independent speaker clustering, speaker recognition experiments were carried out using the Harvard sentence database. Nine male speakers uttered ten different Harvard sentences each. Codebooks were generated from the first five sentences for each speaker using Weighted Likelihood Ratio measure (WLR) through LPC analysis. Using 128 vectors in each codebook, a speaker recognition rate of 98% was attained on the latter five Harvard sentences. Effects of codebook size and input length are also discussed. The above approach based on framewise VQ only utilizes the static distribution of LPC spectra. VQ for multiframe codebooks was used to represent the coarticulation units. The results of speaker recognition experiments based on multi‐frame codebooks will be compared with fixed length VQ approaches.

Comparison of speaker recognition methods using statistical features and dynamic features

This paper describes results of speaker recognition experiments using statistical features and dynamic features of speech spectra extracted from fixed Japanese word utterances. The speech wave is transformed into a set of time functions of log area ratios and a fundamental frequency. In the case of statistical features, a mean value and a standard deviation for each time function and a correlation matrix between these functions are calculated in the voiced portion of each word, and after a feature selection procedure, they are compared with reference features. In the case of dynamic features, the time functions are brought into time registration with reference functions. The results of the experiments show that there is only a slight difference between the recognition accuracies for statistical features and dynamic features over the long term. Since the amount of calculation necessary for recognition using statistical features is only about one-tenth of that for recognition using dynamic features, it is more efficient to use statistical features than dynamic features. When training utterances are recorded over ten months for each customer and spectral equalization is applied, 99.5 percent and 96.3 percent verification accuracies can be obtained for input utterances ten months and five years later, respectively, using statistical features extracted from two words. Combination of dynamic features with statistical features can reduce the error rate to half that obtained with either one alone.

Effects of long-term spectral variability on speaker recognition

One of the most difficult problems in speaker recognition is that the feature parameters frequently vary after a long time interval. We examined this effect on two kinds of speaker recognition; one uses the time pattern of both the fundamental frequency and log-area-ratio parameters and the other uses several kinds of statistical features derived from them. Results of speaker recognition experiments revealed that the long-term variation effects have a great influence on both recognition methods, but are more evident in recognition using statistical parameters. In order to reduce the error rate after a long interval, it is desirable to collect learning samples of each speaker over a long period and measure the weighted distance based on the long-term variability of the feature parameters. When the learning samples are collected over a short period, it is effective to apply spectral equalization using the spectrum averaged over all the voiced portions of the input speech. By this method, an accuracy of 95% can be obtained in speaker verification even after five years using statistical parameters of a spoken word.

Speaker recognition using orthogonal linear prediction

The effectiveness of a set of speaker recognition features is usually characterized in terms of the ratio of the interspeaker variability of the feature to its intraspeaker variability (F-ratio). A recent experiment in speech synthesis [M.R. Sambur, “An Efficient LPC Vocoder,” Bell Syst. Tech. J. (to be published)] has shown that by an appropriate eigenvector analysis, a set of orthogonal parameters can be obtained that is essentially constant across an utterance for a given speaker (i.e., zero intraspeaker variability). If the same eigenvector analysis is applied to the same utterance spoken by another speaker, the resulting values of the orthogonal parameters are, however, different. These orthogonal parameters were therefore examined for their ability to differentiate different speakers. They were formally tested in a speaker recognition experiment involving 21 speakers. The speech data consisted of six repetitions of the same sentence spoken by each speaker on six separate occasions. The identification and verification accuracy of the orthogonal parameters exceeded 99%.