Wav2Vec2-based Automatic Speech Recognition Systems for Filipino Children’s Speech

Developing automatic speech recognition (ASR) systems for Filipino children’s speech remains a significant challenge despite the existing need for automated reading tutors (ARTs) to support early literacy instruction. This paper explores the wav2vec2 adaptation for the Tanglaw corpus, and developed two baseline models: one pre-trained with Filipino adult’s speech, and another on English children’s speech. It also evaluates two data augmentation strategies which are Spectrogram Augmentation (SpecAugment) and MaskCycleGAN-based voice conversion (MaskCycleGAN-VC) to address the low-resource nature of Filipino children’s speech. Across ten model configurations, the best-performing system achieved a Word Error Rate (WER) of 1.80% which substantially outperforms both baseline systems. This result demonstrates an effective approach to developing Filipino children’s ASR systems which could particularly be valuable in resource-limited learning environments.