گوناگوني بين‌گوينده سايشي‌هاي بي‌واك زبان فارسي در سبك‌هاي گفتاري خوانشي و واضح با استفاده از الگوريتم‌هاي يادگيري ماشين

Between-Speaker Variability of Voiceless Fricatives of Persian in Read an‎d Clear Speaking Styles Using Machine Learning Methods

زبان شناسي

This research investigates the discriminative capabilities of Persian voiceless fricatives in speaking style classification an‎d speaker identification tasks. Two distinct sets of acoustic parameters were extracted from voiceless fricative segments: first, spectral moments (including center of gravity, stan‎dard deviation, skewness, an‎d kurtosis), an‎d second, Mel-frequency cepstral coefficients. The classification performance was eva‎luated using three machine learning algorithms: k-nearest neighbor, support vector machine, an‎d ran‎dom forest. For speaking style identification, results demonstrated that the ran‎dom forest algorithm utilizing spectral moment features achieved a maximum accuracy of 61.6%. In contrast, the k-nearest neighbor algorithm employing Mel-frequency cepstral coefficient features yielded superior classification performance with 82.9% accuracy. The combined implementation of spectral moments an‎d Mel-frequency cepstral coefficients maintained this 82.6% classification accuracy when processed by the k-nearest neighbor algorithm. In speaker identification tasks, spectral moment features exhibited limited effectiveness across both speaking styles (clear an‎d read), with the k-nearest neighbor algorithm attaining only 13.7% identification accuracy in the clear speaking style condition. However, implementation of Mel-frequency cepstral coefficient features substantially enhanced system performance, with the k-nearest neighbor algorithm achieving 79.5% identification accuracy in clear speaking style. The combined feature approach showed slightly inferior performance (76.8% accuracy in k-nearest neighbor algorithm) compared to using Mel-frequency cepstral coefficients alone. The analysis revealed approximately 1% higher accuracy for speaker identification in clear speaking style compared to read speaking style. Phoneme-specific eva‎luation indicated that /s/ an‎d /ʃ/ fricatives contained the most speaker-specific information when analyzed individually, while the /x/ fricative demonstrated exceptional performance (90.2% accuracy) under the ran‎dom forest algorithm when using combined features. These findings collectively demonstrate that Mel-frequency cepstral coefficients extracted from Persian voiceless fricatives serve as effective acoustic parameters for both speaking style classification an‎d speaker identification. The results further suggest that spectral moments an‎d Mel-frequency cepstral coefficients provide complementary information for capturing speaker characteristics across different speaking styles.

5

اليسا پلگرينو