Fine-Tuning Wav2Vec2 for Low-Resource Kichwa Automatic Speech Recognition

In recent years, advancements in artificial intelligence (AI) have significantly accelerated the development of natural language processing and automatic speech recognition (ASR) systems for high-resource languages, raising concerns about the marginalization of ancestral and underrepresented languages. In this context, this work explores the fine-tuning of the Wav2Vec 2.0 model, developed by Meta AI, for ASR in Kichwa-a low-resource language spoken in the Ecuadorian Andes. The training process utilized two datasets totaling approximately 8 hours of audio, segmented into clips ranging from 1.5 to 5 seconds, with manually aligned transcriptions created using ELAN software. Fine-tuning was performed using the Connectionist Temporal Classification (CTC) loss function. After multiple experiments, a two-tailed Wilcoxon signed-rank test revealed no statistically significant improvement when applying SpecAugment. The best-performing model, trained without data augmentation, achieved promising results on the test set: a Word Error Rate (WER) of 0.262, a Character Error Rate (CER) of 0.120, and a Match Error Rate (MER) of 0.401. These findings d emonstrate the viability of adapting pre-trained self-supervised models to low-resource settings and underscore the potential of ASR technologies to support greater linguistic inclusivity in artificial intelligence.