Salvatore Casale

A robust voice activity detector for wireless communications using soft computing

Discontinuous transmission based on speech/pause detection represents a valid solution to improve the spectral efficiency of new generation wireless communication systems. In this context, robust voice activity detection (VAD) algorithms are required, as traditional solutions present a high misclassification rate in the presence of the background noise typical of mobile environments. This paper presents a voice detection algorithm which is robust to noisy environments, thanks to a new methodology adopted for the matching process. More specifically, the VAD proposed is based on a pattern recognition approach in which the matching phase is performed by a set of six fuzzy rules, trained by means of a new hybrid learning tool. A series of objective tests performed on a large speech database, varying the signal-to-noise ratio (SNR), the types of background noise, and the input signal level, showed that, as compared with the VAD standardized by ITU-T in Recommendation G.729 annex B, the fuzzy VAD, on average, achieves an improvement in reduction both of the activity factor of about 25% and of the clipping introduced of about 43%. Informal listening tests also confirm an improvement in the perceived speech quality.

Performance evaluation and comparison of G.729/AMR/fuzzy voice activity detectors

The paper proposes a performance evaluation and comparison of G.729, AMR, and fuzzy voice activity detection (FVAD) algorithms. The comparison was made using objective, psychoacoustic, and subjective parameters. A highly varied speech database was also set up to evaluate the extent to which VADs depend on language, the signal-to-noise ratio (SNR), or the power level.

Speech Emotion Classification Using Machine Learning Algorithms

The recognition of emotional states is a relatively new technique in the field of machine learning. The paper presents the study and the performance results of a system for emotion classification using the architecture of a distributed speech recognition system (DSR). The features used were extracted by the front-end ETSI Aurora eXtended of a mobile terminal in compliance with the ETSI ES 202-211 V1.1.1 standard. On the basis of the time trend of these parameters, over 3800 statistical parameters were extracted to characterize semantic units of varying length (sentences and words). Using the WEKA (Waikato Environment for Knowledge Analysis) software the most significant parameters for the classification of emotional states were selected and the results of various classification techniques were analysed. The results, obtained using both the Berlin Database of Emotional Speech (EMO-DB) and the Speech Under Simulated and Actual Stress (SUSAS) corpus, showed that the best performance is achieved using a support vector machine (SVM) trained with the sequential minimal optimization (SMO) algorithm, after normalizing and discretizing the input statistical parameters.

Multistyle classification of speech under stress using feature subset selection based on genetic algorithms

The determination of an emotional state through speech increases the amount of information associated with a speaker. It is therefore important to be able to detect and identify a speakers emotional state or state of stress. Various techniques are used in the literature to classify emotional/stressed states on the basis of speech, often using different speech feature vectors at the same time. This study proposes a new feature vector that will allow better classification of emotional/stressed states. The components of the feature vector are obtained from a feature subset selection procedure based on genetic algorithms. A good discrimination between neutral, angry, loud and Lombard states for the simulated domain of the Speech Under Simulated and Actual Stress (SUSAS) database and between neutral and stressed states for the actual domain of the SUSAS database is obtained.

Performance evaluation and comparison of ITU-T/ETSI voice activity detectors

The paper proposes a performance evaluation and comparison of recent ITU-T and ETSI voice activity detection algorithms. The comparison was made using both objective and psychoacoustic parameters, so as to have reliable judgements that were close to subjective ones. A highly varied speech database was also set up to evaluate the extent to which VAD depend on language, the signal to noise ratio, or the power level.

Multilevel Speech Classification Based on Fuzzy Logic

In the recent generation of v e 9 low bit-rate speech coding schemes, one of the most delicate issues is to adapt the appropriate signal excitation to the LPC filter modeling the vocal tract. The problem essentially consists of the need for a good, efficient speech pame classifier. The paper proposes a new method for multilevel speech classijcation based on Fuzzy Logic. Through simple fuzzy rules, our Fuzty Voicing Detector (FLD) system achieves a sophisticated speech classiification, returning a range of continuous values between the two extreme classes of voiced/unvoiced. As compared with traditional algorithms, the FT/2) correctly classifies typically difficult sound and, on account of its fuzzy nature, maintains good performance even in presence of hackpround noise.

A low‐complexity speech‐pause detection algorithm for communication in noisy environments

The paper presents a new low-complexity algorithm for silence suppression in adverse acoustic environments. The algorithm uses a single time-domain input parameter (signal power) given to a simple matching block. The decision module adapts a series of thresholds depending on the current estimated signal-to-noise-ratio (SNR) of the signal. A series of tests carried out using a large speech database confirm a 10% improvement in pause detection performance as compared with the AMR VAD option 1 recently adopted by ETSI for 3rd-generation mobile systems. Copyright

Robust voiced/unvoiced speech classification using fuzzy rules

The paper presents a robust voiced/unvoiced speech classifier based on fuzzy logic. More specifically, the classification is based on a pattern recognition approach in which the matching phase is performed using a set of 5 fuzzy rules obtained by training. Certain interesting statistical properties of the fuzzy system allow the transition threshold to be adapted to the level of background noise. The results show that the performance of the fuzzy classifier in the presence of various types of background noise is better than that of traditional methods.

Robust phase reversal tone detection using soft computing

The paper presents a simple robust algorithm for the recognition of a 2100 Hz tone with periodic phase reversal and the disabling of an echo canceller based on soft computing. The authors have used a novel tool that is able to extract fuzzy knowledge using a hybrid technique based on genetic algorithms and neural networks. The approach proposed, compared with signal detection solutions existing in literature, is certainly more efficient in terms of robustness to channel noise and can therefore be usefully applied in all cases in which signals are to be detected with very low SNRs.

New results in fuzzy pattern classification of background noise

This paper proposes a background noise classifier based on a new, computationally simple, robust set of acoustic features. Complementary to our previous work (1998), reporting on the first studies carried out by the authors on background noise classification, this paper mainly presents: 1) a criterion to group a large range of environmental noise into a reduced set of classes of noise with similar acoustic characteristics; 2) a larger set of background noise together with a new multilevel classification architecture; and 3) a new set of robust acoustic parameters. We have maintained the pattern recognition approach proposed previously in which the matching phase is performed using a set of trained fuzzy rules. The improved version of the fuzzy noise classifier has been assessed in terms of misclassification percentage and compared with the quadratic Gaussian classifier.