Addition of a new Paper

[hemmetaverse]

Development of a code-switched Hindi-Marathi dataset and transformer-based architecture for enhanced speech recognition using dynamic switching algorithms

Highlights

1. Developed a 450-hour Hindi-Marathi dataset with balanced intra- and inter-sentential code-switching instances.
2. Employed Q-Learning, SARSA, and DQN algorithms to dynamically determine optimal language switch points in speech data.
3. Achieved WER of 0.2800 and CER of 0.2400, surpassing heuristic methods and monolingual baselines for code-switched ASR tasks.
4. Demonstrated that transformer-based ASR models excel at handling code-switching in challenging low-resource scenarios.
5. Conducted extensive hyperparameter tuning, including dropout, learning rates, and regularization, for better ASR models.

https://www.sciencedirect.com/science/article/abs/pii/S0003682X24005590