Monica Feraru

Speech emotion recognition for SROL database using weighted KNN algorithm

In this study, we utilized an improved version of the classical KNN algorithm which associates to each parameter from the features vectors weights according to their performance in the classification process. We obtained the recognition percents of emotions around 65-67%, for the Romanian language, on the SROL database, which are comparable with the results for other languages, with non-professional voice database. This is the first study when the parameters are extracted on the sentence level. Until now, the analysis was made on the phoneme level.

Statistical characteristics of the formants of the Romanian vowels in emotional states

We investigate the influence of the emotions expressed by female and male speakers on the formants of the vowels in the Romanian language. The main conclusions are that the formant F4 is not significantly modified by the expressed emotion; the pitch is significantly increased for joy in all vowels; the pitch is slightly decreased under sadness, and that the change in statistic distribution is quite large in all cases (vowels and states).

Using the Lyapunov exponent from cepstral coefficients for automatic emotion recognition

The main goal of this paper is to establish the relevance of nonlinear parameters (Lyapunov exponents) in the automatic classification of emotions, for the Romanian language. The Largest Lyapunov Exponent - LLE was computed for the MFCC mel frequency cepstral coefficients and the LPCC linear prediction cepstral coefficients. The Support Vector Machine - SVM classifier provides better results than Weighted K-Nearest Neighbors - WKNN classifier in emotion recognition for feature vectors that contains LLE (around 75%). The best recognized by using SVM classifier was the neutral tone, followed by the sadness, fury and the weakest recognized was the joy. For features vectors which include LLE the best results was obtained in combination with LAR - Log Area Ratio coefficients, respectively PARCOR - partial correlation coefficients.

Emotion recognition using Lyapunov exponent of the Mel-frequency energy bands

This paper presents a method for emotion recognition by using LLE - Largest Lyapunov exponent of the Mel-frequency energy bands for the Romanian language. The emotion recognition for features vectors that contains LLE is better using Support Vector Machine - SVM classifier (76.4%) than Weighted K-Nearest Neighbors - WKNN classifier (72.8%). The most efficient combination was LLE with LPC - linear predictive coefficients, respectively with PARCOR - partial correlation coefficients. The best emotion recognized by using WKNN classifier is the joy state (70-80%) and the least recognized is neutral tone.

Integrated system for prosodic features detection from speech

The paper presents the instruments implemented in SROL (Voiced Sound of Romanian Project) used for the extraction of the fundamental frequency F0 and of the formants F1-F4. In order to have a better detection of the prosodic features, we put together four methods: autocorrelation method, cepstral methods, AMDF (average magnitude difference function), HPS (harmonic product spectrum). The final value of pitch (F0) is established after an integration of the partial results of each method depending on their performance. For the formants detection we used a fuzzy method of spectrum concatenation.

Assessing the quality of voice synthesizers

We address the voice quality assessment by focusing on the harmonic and temporal aspects that are less treated in the literature. The main focus is on the influence on synthetic voice quality of the higher formants. We view the higher formants role as primarily mediated by the ratios of their frequencies and that of the pitch. We present a comparative study between human and synthetic voice in this respect. The presented analysis involves, beyond the determinations of the fundamental frequency and formants, the determination of the durations of vowels. Recognizing that a single quality criterion could not give a sufficient understanding of the voice quality process, we suggest a set of criteria regarding the discussed aspects of the voice and introduce a set of parameters to characterize the quality.

Analysis of vowel triangle variation for the Romanian language related to emotional states

The vowel triangle for Romanian language, for normal tone and for three emotional states are determined and statistically analyzed.