Abstract:
The development of voice recognition systems tailored to vernacular dialects holds
transformative potential for enhancing accessibility and inclusivity in technology. This thesis
focuses on creating a voice recognition model specifically designed for vernacular Gujarati
dialects, addressing the unique linguistic and phonetic challenges inherent in regional
variations of the language.
The key part of this research was to gather a diverse and representative spoken Gujarati corpora
sourced via varied public repositories, which includes radio broadcast, interview, folk song,
community recording and public availability speech corpora. This dataset includes a variety of
dialectal variation in phonology, syntax and usage to guarantee robustness and inclusivity to
the development of the models.
A dialect-specific recognition system using advanced techniques in voice recognition system,
including deep learning architectures the proposed framework and model was developed. The
model is further enriched with dialectal linguistic features integrated to its architecture,
phoneme based pretraining to increase recognition accuracy, and transfer learning to adapt
general speech recognition systems to dialect specific nuances.
The model was evaluated and found to achieve substantial improvement in phoneme
recognition accuracy over baseline systems. The results show that modeling context-aware,
high quality, diverse datasets are crucial to vernacular speech recognition. The system
developed is there to provide practical applications for voice enabled user interface, digital
accessibility and protection of linguistic diversity more specific examples of such languages
which are least represented.
This work contributes to the emerging area of regional language processing with an end-to-end
framework that can be used for future work on low-resource languages and dialects and to build
inclusive, ubiquitous and accessible technology solutions in multilingual communities.
