Transfer learning through perturbation-based in-domain spectrogram augmentation for adult speech recognition

The development of numerous frameworks and pedagogical practices has significantly improved the performance of deep learning-based speech recognition systems in recent years. The task of developing automatic speech recognition (ASR) in indigenous languages becomes enormously complex due to the wide range of auditory and linguistic components due to a lack of speech and text data, which has a significant impact on the ASR system's performance. The main purpose of the research is to effectively use in-domain data augmentation methods and techniques to resolve the challenges of data scarcity, resulting in an increased neural network consistency. This research further goes into more detail about how to create synthetic datasets via pooled augmentation methodologies in conjunction with transfer learning techniques, primarily spectrogram augmentation. Initially, the richness of the signal has been improved through the process of deformation of the time and/or the frequency axis. The time-warping aims to deform the signal's envelope, whereas frequency-warping alters spectral content. Second, the raw signal is examined using audio-level speech perturbation methods such as speed and vocal tract length perturbation. These methods are shown to be effective in addressing the issue of data scarcity while having a low implementation cost, making them simple to implement. Nevertheless, these methods have the effect of effectively increasing the dataset size because multiple versions of a single input are fed into the network during training, likely to result in overfitting. Consequently, an effort has been made to solve the problem of data overfitting by integrating two-level augmentation procedures via pooling of prosody/spectrogram modified and original speech signals using transfer learning techniques. Finally, the adult ASR system was experimented on using deep neural network (DNN) with concatenated feature analysis employing Mel-frequency cepstral coefficients (MFCC), pitch features, and the normalization technique of Vocal Tract Length Normalization (VTLN) on pooled Punjabi datasets, yielding a relative improvement of 41.16 percent in comparison with the baseline system.