R
Rosie
Text
Audio Data Augmentation for Speech Recognition in Low-Resource Languages
How an end-to-end framework fine-tunes the Wav2Vec2-xlsr-53 acoustic model on augmented audio to overcome data scarcity and heavy dialectal variation in low-resource languages. The learner sees three waveform augmentations - band-stop filtering (simulating accents), additive Gaussian noise (robustness to noisy input), and pitch shifting (synthesizing new speakers) - combined to enlarge tiny Common Voice training sets for Arabic, Russian, and Portuguese, cutting Word Error Rate by ~34% and Character Error Rate by ~50% versus plain Wav2Vec2 and Whisper baselines.