Optimizing Low-Resource Language Translation and Speech Recognition in AI Multi-lingual Virtual Assistants Using Domain Adversarial Neural Network

Artificial intelligence (AI)-powered multi-lingual virtual assistants enable seamless communication across diverse languages. However, effective translation and speech recognition for low-resource languages (LRL) remains challenging due to limited training data and linguistic diversity. Current translation and speech recognition systems often underperform in the LRL environment, mainly due to inadequate domain adaptation and scarcity of labeled datasets. This leads to high error rates and reduced effectiveness of AI assistants in multi-lingual environments. To address these challenges, this study introduces the Domain Adversarial Neural Network-Driven Multi-lingual Adaptation Framework (DANN-MAF). The framework leverages adversarial neural networks to enhance domain adaptation, combining semi-supervised learning and distribution alignment techniques to bridge the performance gap between high-resource and LRLs. Experimental results on standard multi-lingual datasets show that DANN-MAF significantly improves translation accuracy and speech recognition performance. Metrics such as BLEU score and Word Error Rate (WER) indicate superior performance compared to baseline models, particularly in underrepresented language domains. The proposed DANN-MAF framework demonstrates the effectiveness of AdaMatch-based domain adaptation in optimizing multi-lingual AI assistants. It offers a scalable solution for enhancing LRL capabilities and fostering inclusive and equitable access to language technologies.

Hali tarjima qilinmagan