We propose a novel method of measuring the similarity between two or more speech utterances for speaker clustering, based on probability theory and factor analysis. The similarity function is formulated as the probability that the utterances originated from the same speaker, and uses statistical eigenvoice and eigenchannel models to incorporate physical knowledge of interspeaker and intraspeaker variabilities, allowing the similarity function to be trainable and robust. The comparison function can be efficiently computed using a compact set of sufficient statistics for each speech utterance, allowing the acoustic features to be discarded. We begin using only eigenvoices, and then show how the eigenchannels can be incorporated into the equation to result in an identical form but with a different set of sufficient statistics. We test the proposed model in a speaker clustering task using the CALLHOME telephone conversation corpus and show that it performs better than two other well-known similarity measures: the Cross-Likelihood Ratio (CLR) and Generalized Likelihood Ratio (GLR).

We propose a novel utterance comparison model based on probability theory and factor analysis that computes the likelihood of two speech utterances originating from die same speaker. The model depends only on a set of statistics extracted from each utterance and can efficiently compare utterances using these statistics without requiring die indefinite storage of speech features. We apply the model as a distance metric for speaker clustering in die CALLHOME telephone conversation corpus to achieve competitive results compared to three other known similarity measures: the Generalized Likelihood Ratio, Cross-Likelihood Ratio, and eigenvoice distance.

Speech activity detection (SAD) is a key objective in speech-related technologies. In this work, an enhanced version of the training stage of a SAD system based on a support vector machine (SVM) classifier is presented, and its performance is tested with the RT05 and RT06 evaluation tasks. A fast algorithm of data reduction based on proximal SVM has been developed and, furthermore, the specific characteristics of the metric used in the NIST SAD evaluation have been taken into account during training. Tested with the RT06 data, the resulting SVM SAD system has shown better scores than the best GMM-based system developed by the authors and submitted to the past RT06 evaluation.

This work aims at gaining an insight into the mean and variance normalization technique (MVN), which is commonly used to increase the robustness of speech recognition features. Several versions of MVN are empirically investigated, and the factors affecting their performance are considered. The reported experimental work with real-world speech data (Speecon) particularly focuses on the recursive updating of MVN parameters, paying attention to the involved algorithmical delay. First, we propose a decoupling of the look-ahead factor (which determines the delay) and the initial estimation of mean and variance, and show that the latter is a key factor for the recognition performance. Then, several kinds of initial estimations that make sense in different application environments are tested, and their performance is compared.

Reliable measures of speaker positions are needed for computational perception of human activities taking place in a smartroom environment. In this work, it is described the development process and the experiments conducted in the design and implementation of an Audio Person Tracking system for smart-room environments. The proposed system is based on the SRP-PHAT algorithm, as it is known to perform robustly in most environmental conditions. Novelties proposed are aimed to enhance the accuracy of the system independently on the application scenario and to reduce the computational complexity. Index Terms: sound localization, source tracking, microphone arrays.

Speech detection becomes more complicated when performed in noisy and reverberant environments like e.g. smart rooms. In this work, we design a robust speech activity detection (SAD) algorithm and we evaluate it on distant microphone signals acquired in a smart room-like environment. The algorithm is based on a measure obtained from applying linear discriminant analysis (LDA) on frequency filtering (FF) features. With a time sequence of this measure, a decision tree based speech/non-speech classifier is trained. The proposed SAD system is evaluated together with other SAD systems (GSM SAD and ETSI advanced front-end standard SAD) using a set of general SAD metrics as well as using the ASR accuracy as a metric. The proposed SAD algorithm shows better average results than the other tested SAD systems for both the set of general SAD metrics and the ASR performance.

To realize the long-term goal of ubiquitous computing, technological advances in multi-channel acoustic analysis are needed in order to solve several basic problems, including speaker localization and tracking, speech activity detection (SAD) and distant-talking automatic speech recognition (ASR). The European Commission integrated project CHIL, “ Computers in the Human Interaction Loop”, aims to make significant advances in these three technologies. In this work, we report the results of our initial automatic source localization, speech activity detection, and speech recognition experiments on the CHIL seminar corpus, which is comprised of spontaneous speech collected by both near- and far-field microphones. In addition to the audio sensors, the seminars were also recorded by calibrated video cameras. This simultaneous audio-visual data capture enables the realistic evaluation of component technologies as was never possible with earlier data bases.

Reliable measures of speaker positions are needed for computational perception of human activities taking place in a smart-room environment. In this work, we investigate the effect of talkers head orientation on the accuracy of acoustical source localization techniques and its relation with the talker directivity pattern and room reverberation. Two different representative speaker localization techniques are assessed, steered response power and a crossing lines based method, in both cases on the basis of the estimated delays between pairs of microphones with the GCC-PHAT algorithm. A small database has been collected at the UPC’s smart room for evaluation. The results show how the localization error heavily depends on the head orientation, and also the fact that the space exploration based technique is much more robust to head orientation changes than the crossing lines technique, due to the way the contributions from the various microphones are combined.

In this paper, we aim to improve the classification of human non-speech sounds produced in a meeting-room environment by using concepts and tools from the fuzzy theory. Starting with an SVM-based baseline system, firstly a reduction of the number of features with the fuzzy measure is shown. And, secondly, a noticeable improvement of the classification performance is reported by combining the outputs of two classification systems with the fuzzy integral.

Wideband audio signal will be commonly used in the near future telecommunication applications. In quiet environments, the speech recognition performance increases when using a wideband signal instead of a narrowband signal. For practical ASR systems, however, we are interested in whether we can benefit from a wideband signal when recognizing noisy speech. The wideband speech signal, with respect to the narrowband speech signal, contains high frequency components, which usually have low intensity and, thus, are vulnerable to noise distortions. The robustness of the wideband feature set may be worse than that of the narrowband feature set, if the added high-frequency components are distorted. We investigate whether the addition of information from high frequencies into the ASR feature set can improve the recognition performance of noisy speech. The differences between low- and high-frequency parts of the wideband speech spectrum suggest a separate processing of these two parts. We propose an algorithm in which the separate processing scheme permits us to reuse the noise robust front-end originally designed for a narrowband signal. A low complexity processing is designed for high frequency components, which usually bear less information. The high-frequency information is added to the narrowband speech features in a form of de-noised filter-bank energies. The energies are appended after or before computing the cepstral features. In the best case, we obtained 13.96% average relative improvement when recognizing wideband noisy speech with respect to the narrowband noisy speech performance. The proposed algorithm is part of the recently adopted ETSI standard for the advanced frontend for distributed speech recognition.

We introduce a new concept in advancing the noise robustness of a speech recognition front-end. The presented method, called SNR-dependent waveform processing (SWP), exploits SNR variability within a speech period for enhancing the high SNR period portion and attenuating the low SNR period portion in the waveform time domain. In this way, the overall SNR of noisy-speech is increased, and at the same time, the periodicity of voiced speech is enhanced. This approach differs significantly from the well-known speech enhancement techniques, which are mostly frequency domain based, and we use it in this work as a complementary technique to them. In tests with SWP, we present significant clean and noisy speech recognition performance gains using the AURORA 2 database and recognition system as defined by ETSI for the robust frontend standardization process. Moreover, the presented algorithm is very simple and it is attractive also in terms of computational load.

This paper presents the robust front-end algorithm that was submitted by Motorola to the ETSI STQ-Aurora DSR working group as a proposal for the Advanced DSR front-end in January 2001. The algorithm consists of a two-stage melwarped Wiener filter, a waveform processor, a channelnormalized mel-frequency cepstral calculation and a subsystem of post-cepstral processing according to the reliability of mel-spectral components, etc. The output of this algorithm, a set of Mel-Frequency Cepstral Coefficients (MFCC), is compressed and encoded based on ETSI ES 201 108 standard; and then it is transmitted at 4800 bps. Compared to ETSI standard MFCC front-end, the proposed algorithm delivers an improvement of 42.64% in performance on the Aurora 2 database, which is required by this Eurospeech Special Event. With a very simple frame deletion algorithm based on a Voice-Onset Detection (VOD), the improvements were significantly boosted to 47.58% on the same database. In this paper we also give further insights about the proposal by providing performances and analyses with the Aurora SpeechDat-Car databases.

Speech recognition systems that operate in real world environments have to be robust against additive noises. In this work, a technique that uses a voicing-dependent exponent in the computation of the filter-bank parameters to improve their robustness to additive noise is presented. Speech recognition experiments with the Aurora 1.0 database and recognition setup are reported to show the potential of this new technique.

In current speech recognition systems, speech is represented by a 2-D sequence of parameters that model the temporal evolution of the spectral envelope of speech. Linear transformation or filtering along both time and frequency axes of that 2-D sequence are used to enhance the discriminative ability and robustness of speech parameters in the HMM pattern-matching formalism. In this paper, we compared two recently reported approaches which operate on the sequence of logarithmically compressed mel-scaled filter-bank energies: the first approach - TIFFING (TIme and Frequency FilterING) - applies FIR filters to that 2-D sequence along both axes, while the second one - CTM (Cepstral Time Matrix) - uses the DCT to compute a set of parameters in the 2-D transformed domain. They are compared in several ways: (1) analytically, using Fourier transformation, (2) statistically and (3) performing recognition tests with clean and noisy speech.