Multilayer Perceptron Classification for Multilingual Speech Detection

Automatically detecting the language spoken in audio recordings, also known as language detection, is needed for several applications like speech recognition, cross-cultural communication, and multilingual translation. Standard speech detection techniques that can handle the processing of speech in specific languages are somehow limited which discriminates the general phonetic information. This research presented a comprehensive study on the advancement of a language detection system that uses a machine learning model to identify speech utterances from various languages and by various speakers. The effectiveness of classifiers was explored through the model that accurately distinguishes different languages based on extracted acoustic attributes from audio signals. The voice source and low-level features in terms of frequency cepstral coefficients like Mel, Gammatone and Power Normalized Cepstral Coefficient of labelled data were extracted. The perceptual analysis was made at different frequency bands while defining the auditory model. These features presented the frequency scale of critical bands like Mel scale and Bark scale etc. that characterized the multilingual speech from the corpus of around 10 Indian languages and were modelled using a Multilayer Perceptron Classifier. This demonstrated an improved efficiency of 92 % with a combined classifier compared with a stand-alone approach when tested on static and real-time data. Thus, this proposed approach will have a reasonable impact on speech translation and many applications and can make use of language as more prevalent data for the machine to build and solve language identification tasks.