Noisy Student Teacher Training with Self Supervised Learning for Children ASR

Automatic Speech Recognition (ASR) is a fast-growing field, where reliable systems are made for high resource languages and for adult's speech. However, performance of such ASR system is inefficient for children speech, due to numerous acoustic variability in children speech and scarcity of resources. In this paper, we propose to use the unlabeled data extensively to develop ASR system for low resourced children speech. State-of-the-art wav2vec 2.0 is the baseline ASR technique used here. The baseline's performance is further enhanced with the intuition of Noisy Student Teacher (NST) learning. The proposed technique is not only limited to introducing the use of soft labels (i.e., word-level transcription) of unlabeled data, but also adapts the learning of teacher model or preceding student model, which results in reduction of the redundant training significantly. To that effect, a detailed analysis is reported in this paper, as there is a difference in teacher and student learning. In ASR experiments, character-level tokenization was used and hence, Connectionist Temporal Classification (CTC) loss was used for fine-tuning. Due to computational limitations, experiments are performed with approximately 12 hours of training, and 5 hours of development and test data was used from standard My Science Tutor (My-ST) corpus. The baseline wav2vec 2.0 achieves 34% WER, while relatively 10% of performance was improved using the proposed approach. Further, the analysis of performance loss and effect of language model is discussed in details.