Exploring the Role of Data Augmentation and FDLP Features on E2E Children's ASR Under Low Resource Conditions

The work presented in this paper aims at enhancing the performance of end-to-end (E2E) speech recognition task for children's data under low resource conditions. Since training a robust E2E automatic speech recognition (ASR) system requires a huge amount of data, extremely poor recognition performances are observed when a limited amount of children's speech is used. To overcome the problems arising from data scarcity, we have resorted to several out-of-domain and in-domain data augmentation techniques. For out-of-domain data augmentation, we have explicitly modified adults' speech to render it acoustically similar to children's speech before pooling into training. On the other hand, in the case of in-domain data augmentation, we have slightly modified the pitch and duration of children's speech in order to create more data capturing greater diversity. In addition to that, we have also explored the efficacy of frequency domain linear prediction (FDLP) technique for front-end speech parameterization. The application of linear prediction on the frequency domain representation of speech signal helps to effectively capture the temporal envelope during front-end feature extraction. In our experimental explorations, the use of proposed data augmentation in combination of FDLP features has shown a relative improvement in character error rate by 92.9% over the baseline ASR system.