Building spoken dialect identification system in low-resource conditions

Abstract Spoken dialect identification (DID) refers to the automatic identification of dialects in a speech sample. Since different dialects of a given language have very high similarity in terms of vocabulary, grammar, pronunciation, etc., accurate identification of the dialect is very challenging. In addition to this, lack of sufficient training data, which is common in low-resource languages, further complicates the issue. Due to these reasons, state-of-the-art DID systems often perform unsatisfactorily. Using a feature representation of the speech that can efficiently encode DID-specific contents in the input even in low-resource conditions, and using a suitable model architecture that can efficiently capture the DID-specific contents in the given input feature representation, including subtle differences between the dialects, can help address this issue. Motivated by this, in this work, we explore the usage of different frame-level and segment-level features, and explore different model architectures for accurate identification of the dialect. Specifically, apart from commonly used Mel-frequency cepstral coefficients (MFCC) features, we explore the usage of bottleneck features and wav2vec2.0 features (both base and large variants) obtained using pre-trained networks as representation of the speech. Following this, we experiment with state-of-the-art x-vector, BLSTM-based-u-vector, and transformer-based-u-vector which uses attention in a hierarchical manner. This helps to efficiently model the DID-specific contents in the input representation of speech. Experiments conducted on Kannada, Konkani, Tamil and Marathi, which are some low-resource languages of India, indicate that wav2vec2.0 large features, when combined with transformer-based model, are better as compared to other combinations of features and model architectures and provided around 10% improvement compared to baselines in all cases.