Lambdacism Detection in L2 English Speech Using Spectrogram Features and Machine Learning Techniques

This study investigates the use of machine learning and feature extraction through spectrograms to identify pronunciation errors, particularly lambdacism, in the word “Alive” by native Igbo speakers. Lambdacism leads to an “r” substitution for the “l” sound, and with words like “arrive”, “alive” becoming prevalent in speech, it diminishes clarity and understanding. This work processes audio samples of right and wrong phonemes using DSP and phoneme transcription. Essential features are extracted through STFT, MFCCs, Gammatone, andcochleagram spectrograms. These features are further reduced for visualization and processed using a One-Class Support Vector Machine (SVM) that trains on the distribution of correctly pronounced words and detects mispronunciations as outliers. Validationof mispronunciation using the correct pronunciation is further strengthened by DTW distance calculations. Spectrogram analysis shows that the correct samples have clearer frequency structures, while the incorrect samples are filled with noise and energy incoherence. One-Class SVM outperformed other models of anomaly detection,like Isolation Forest and Local Outlier Factor,by demonstrating high sensitivity in detecting minute errors at the phoneme level. An evaluation based on the F1 score has shownthat the model has an equilibrium of precision and recall. Not only does this approach demonstrate the phonetic effects of lambdacism, but it also shows a strong system for the automated detection of phonetic errors in pronunciation. The results provide supportfor further development as components of a speech training system that is interactive and adaptable, that caters to users with phonological disorders and difficulties with second language pronunciation.