Voice Conversion Based Data Augmentation Using CycleGAN for Children's ASR

Extensive use of voice assistants by children in their day-to-day life activities demand for better performance of Automatic Speech Recognition (ASR) for children's speech. The recent advancements in ASR perform better for adult speech. However, due to acoustic mismatch (in particular, higher
$F_{0}$
and thus, poor spectral resolution, and less availability of children data), it remains a challenge to improve the performance of children's ASR. In this context, a shared task on ASR for children's speech 2021 was organized during INTERSPEECH 2021. In this paper, the voice conversion-based data augmentation technique using a cycle-consistent generative adversarial network (CycleGAN) is investigated for closed track in the challenge. Here, CycleGAN training is exploited for children-to-children voice conversion. Training of CycleGAN and ASR experiments are performed on the speech data provided during the challenge. In our experiments, CycleGAN-based data augmentation showed a relative improvement of 9% and 3% in WER compared to the baseline system on the Dev and Eval sets, respectively. This work may find its significance in ASR for low resource languages, where Librispeech or any other external dataset is not an option.