Low-Resource Speech Recognition

Low-resource speech recognition studies how to transcribe languages, dialects, and specialized speech when labeled audio is scarce, enabling wider access to speech technology and supporting language documentation. Common approaches adapt pretrained models such as Whisper and wav2vec, transfer knowledge across languages, use self-supervised learning and pseudo-labels, or expand training data with augmentation and speech synthesis. Evaluation increasingly includes spontaneous, noisy, code-switched, and domain-specific speech, with attention to scripts, dialects, and error types beyond aggregate word error rate. Large multilingual models provide useful starting points, but zero-shot performance and cross-lingual transfer remain unreliable; data quality, linguistic coverage, and consistent evaluation often matter as much as model scale.