نظام آلي لتمييز اللكنات

الملخص EN

We covered in this research the building of an automatic accent-discrimination (Accent Recognition) system, and its application on a closed group of English speakers which contains people of Arab descent, Americans and people from the UK.

The system classify a range of different accents among the group.

About sixty audio clips were recorded and used in the research.

The clips were equally divided into two sets: a training set and a test set.

The system labels each of the test clips based on the characteristics (traits) abstracted from it that match the voice traits that were abstracted from the training set and stored in the database.

Hence, the research is divided into two phases: (1) The extraction of important characteristics from the 30 training clips and storage of such data in the database, and (2) the extraction and matching of characteristics from clips (of the test set) to the stored data in the database to recognize and classify the accent In the extraction phase, we used the Mel frequency Cepstral coefficient, which is commonly used by researchers in the field for identifying special traits in voices.

In the classification stage we had experimented with three voice-discrimination/recognition models including 2 statistical-based methods: Linear discriminant analysis and the Correlation distance, and the Dynamic Time Wrapping method which is based on distance measurement.

The ability of each of these three methods to distinguish accents depends on the amount of training data and (some of which depend heavily on) the extent of the correlation among data elements of the same class In the extraction phase, we used the Mel frequency Cepstral coefficient, which is commonly used by researchers in the field for identifying special traits in voices.

In the classification stage we had experimented with three voice-discrimination/recognition models including 2 statistical-based methods: Linear discriminant analysis and the Correlation distance, and the Dynamic Time Wrapping method which is based on distance measurement.

The ability of each of these three methods to distinguish accents depends on the amount of training data and (some of which depend heavily on) the extent of the correlation among data elements of the same class