Algorithms for interpretation of prosodic features in low-bitrate speech processing

We study the language identification problem using prosodic features. Prosodic features such as melody, rhythm, timbre and others are difficult to formalize mathematically. Two algorithms for a complex description of prosodic features are proposed in the paper. The first is based on the broad phonetic categories, and the second is based on the cross-correlation of the speech melody and the short-term energy sequence. The fundamental frequency was estimated by MELP algorithm. The performance of the proposed algorithms was evaluated experimentally on a database of speech recordings obtained from Internet and therefore encoded by low-bitrate vocoders. The database includes ten different languages. The proposed algorithms provide a feature description and a multi-layer neural network was used as a language classifier. Both algorithms show satisfactory classification performance, but the broad phonetic categories approach performs slightly better than the cross-correlation function. These algorithms can be applied to a speech signal processed by low-bitrate vocoders without decoding to the original signal.