Documentation
See my page on SpeechMaster (https://thoughtpocket.net/speechmaster) for more info on the software BengaliWave2sec is based on
01
Training Environment
Utilizing the computational resources of multiple Colab Pro+ accounts, I optimized my training process to accommodate the demands of the Bengali speech recognition task. Despite equipment and computing power constraints, I fine-tuned my models for competitive performance.
02
Novel Techniques
In my pursuit of refining model robustness, I employed a combination of data augmentation techniques. Alongside conventional methods like Spectrogram Augmentation and Time Warping, I introduced novel approaches such as SpecAugment++ and Random Frequency Masking to enhance adaptability to diverse speech patterns.
03
Model Architecture
My approach focused on implementing a new deep learning architecture, named BengaliWav2Vec. Drawing inspiration from successful Wav2Vec models in various languages, BengaliWav2Vec was customized to handle the linguistic intricacies and prosodic variations found in Bengali speech.
04
Training Strategy
I implemented a training strategy, unfreezing model layers gradually and adjusting learning rates dynamically. This iterative process fine-tuned parameters effectively, resulting in performance gains across validation and test datasets.
Featured Video
Vivamus id euismod massa. Suspendisse orci est, cursus quis efficitur quis, porttitor sit amet justo. Sed ac orci felis. Phasellus commodo arcu placerat purus vulputate fringilla. Vivamus rutrum eu tortor non ultrices. Proin posuere porta ligula, id fermentum nisl vestibulum id. Nunc id neque ac magna fringilla ultrices vel quis sem. Fusce a sem vel nunc sollicitudin consectetur. Suspendisse gravida tempor diam, eget imperdiet dui dictum a. Aliquam hendrerit aliquam vehicula. Integer feugiat arcu sapien, in facilisis lacus rutrum eu. Pellentesque eget fringilla odio, et vestibulum turpis. Proin sit amet venenatis nisi.