Enhancing Phoneme Recognition in the Bengali Language Through Fine-Tuning of Multilingual Model

Phoneme recognition is essential in speech recognition systems, allowing for precise transcription of spoken language into text. This technology underpins virtual assistants, voice search, and automated transcription services. It is also vital in language learning applications, where it helps with pronunciation training and accent reduction by offering real-time feedback on spoken language accuracy. Phonemes are the distinct sounds that create meaningful contrasts within a specific language, while phones refer to the actual sounds produced, which are generally language-independent. Allophones are the set of phones that correspond to a particular phoneme. Phoneme recognition of low resource languages using Automatic Speech Recognition (ASR) model produce suboptimal results, hence to improve the result, we can use pre-trained model with fine-tuning. In this work, we are utilizing the multilingual pre-trained model by fine-tuning it for the Bengali language. A multilingual pre-trained model called Allosaurus makes predictions over a common phone inventory before using an allophone layer to map onto a particular phoneme. For fine-tuning, we require phoneme transcribed dataset of Bengali language which is not available. Here we have used one of the text to phoneme model to get phoneme transcribed data. Through our proposed approach, we are able achieve to significant improvements in the phoneme recognition of Bengali language that we can see as the Phoneme Error Rate (PER) of our test file reduces by 95.65% compared to the model without fine-tuning.