Hybrid Feature Based Spoken Language Identification on a 62-Language Dataset Using Machine Learning
CIRCUITS, SYSTEMS, AND SIGNAL PROCESSING, cilt.2026, ss.1-23, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 2026
- Basım Tarihi: 2026
- Doi Numarası: 10.1007/s00034-026-03704-1
- Dergi Adı: CIRCUITS, SYSTEMS, AND SIGNAL PROCESSING
- Derginin Tarandığı İndeksler: Academic Search Ultimate (EBSCO), Engineering Source (EBSCO), Scopus, Materials Science & Engineering Collection (ProQuest), Pharma Collection (ProQuest), Technology Collection (ProQuest), Aerospace Database, Science Citation Index Expanded (SCI-EXPANDED), Compendex, zbMATH
- Sayfa Sayıları: ss.1-23
- Kayseri Üniversitesi Adresli: Evet
Özet
Determining the language spoken by an unknown speaker from a short audio recording requires extracting acoustic features that distinguish between languages and analysing them with appropriate classification methods. In this study, MFCC (Mel-Frequency Cepstral Coefficients) and delta coefficients and mel-spectrogram-based features were obtained from audio recordings and this hybrid feature set was classified using various machine learning models. For the classification phase, 62 of the most prevalent languages were selected from the 136 languages in the Common Voice database, and a dataset consisting of a total of 62,000 voice recordings was created by using 1000 samples of a maximum of 5 s in a balanced manner for each class. This structure, contrary to many studies in the literature, has made the model more challenging in terms of high-class numbers and language diversity. Various machine learning algorithms were tested on the generated hybrid feature sets and the highest accuracy value was obtained by the KNN (K-Nearest Neighbors) algorithm (k = 1) with 85.82%. In addition, in a tenfold cross-validation analysis performed on the same data set, it was observed that the KNN algorithm exhibited one of the most consistent prediction performances with an average accuracy of 98%. These results suggest that the model offers an effective approach for the multi-class language recognition problem in terms of both overall accuracy and cross-validation.