Acoustic analysis of selected homographs for speech recognition systems

This paper presents an acoustic analysis of selected homographs in the context of automatic speech recognition (ASR) systems. The study focuses on the Polish words “Dania” (eng. Denmark) and “dania” (eng. meals), which, despite identical spelling, differ subtly in pronunciation. These differences pose challenges for ASR systems, especially when context is unavailable. The methodology includes spectrograms analysis MFCC (Mel-Frequency Cepstral Coefficients) extraction, and classification using a Support Vector Machine (SVM) algorithm. A custom audio database was created using recordings from ten speakers, followed by manual segmentation and normalization of samples. Spectrograms and formant trajectories were analyzed to identify phonetic distinctions, particularly the presence of the semi-vowel [j] in “Dania”. A subjective listening test involving 27 participants was conducted to assess human recognition accuracy. Results showed an average recognition rate of 58%, indicating significant ambiguity. In contrast, the machine learning model achieved up to 79% accuracy with randomly stratified data and 75% accuracy when tested on the same samples used in the subjective test. The findings suggest that MFCC-based classification combined with SVM is a promising approach for distinguishing homographs in speech, outperforming human listeners in controlled conditions. Limitations include the small dataset and variability in speaker articulation. The study highlights the importance of phonetic exception handling in ASR systems and proposes extending the method to other homographic pairs.