Fine-Tuning Neural ASR Models for Low-Resource Kazakh Children’s Speech with Preprocessing Enhancements

Zhansaya Duisenbekkyzy, Diana Rakhimova, Akgul Yessirkepova, Eşref Adalı

2025

Data confidence 70 medium

How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

This article presents a comprehensive study of automatic speech recognition (ASR) for Kazakh children's speech, with a particular focus on the impact of signal preprocessing techniques in a low-resource linguistic environment. Although commercial ASR systems such as Whisper and DeepSpeech have shown recent progress, their accuracy on children's speech remains limited. We propose a robust ASR pipeline that incorporates Wiener filtering, denoising autoencoders, MFCC/log-Mel features, and transformer-based models (Whisper, Conformer). A novel Kazakh child speech corpus was collected from children aged 2 to 9 via Telegram bots, spontaneous recordings, and curated phrases. We evaluated four neural ASR models under noisy and clean conditions. Our experiments demonstrate that preprocessing significantly improves performance: Whisper achieved WER = 15.4%, CER = 12.3%, Accuracy = 75.2%, BLEU = 0.90. The proposed pipeline is effective for improving ASR accuracy for spontaneous and emotionally variable child speech in low-resource environments.