Speech Representation System For The Karakalpak Automatic Speech Recognition

Abstract Recently, automatic speech recognition (ASR) has moved from deeplearning systems toward end-to-end (E2E) neural architectures. E2E models obtain strong results for high-resource languages, but training reliable systems remains difficult for low-resource languages because labeled speech is expensive to collect and verify. Self-supervised learning (SSL) has reduced this dependency by learning speech representations from large unlabeled corpora and then adapting them to a target language with a smaller amount of transcribed data. In this paper, we present a Karakalpak ASR system based on multilingual speech representations. The work follows the wav2vec 2.0 XLS-R approach, and the Karakalpak-specific acoustic model is trained and fine-tuned by the authors using the Karakalpak Speech Corpus (KSC). KSC is a publicly available benchmark speech-to-text dataset collected by the authors and released through Mendeley Data; it contains 50 hours of predominantly read speech from native Karakalpak speakers. The ASR model is fine-tuned with various KSC data setups, and the effect of the amount of labeled speech on recognition quality is discussed. In addition, word-level n-gram language models are used to improve decoding under low-resource conditions. Recognition error analysis is provided in terms of substitution, insertion, and deletion errors. The 50-hour KSC-trained wav2vec2 XLS-R model gives 21.10\% WER and 4.32\% CER on a held-out test set, and the experimental design treats this result as a reference point for further benchmarking.