CSAF: Child-like Speech Augmentation Framework through Adult Speech Data Modulation

• JunHwi Moon,
• OkJoo Choi,
• Sarah Choi,
• Keon Chul Park,
• JeongRok Lee,
• Wonsun Shin,

Vol. 19, No. 11, November 30, 2025
10.3837/tiis.2025.11.019, Download Paper (Free):

• adult speech data modulation
• ASR
• automatic speech recognition
• child speech recognition
• voice data augmentation

Abstract

The lack of child speech data significantly degrades the performance of commercial automatic speech recognition (ASR) systems. However, it remains challenging to acquire sufficient child speech data because of difficulties in controlling recording conditions and limited cooperation of children during data collection. To address these limitations, this study proposes a data augmentation method to enhance speech resources for developing ASR systems targeted at children aged 4 to 7. Several acoustic modulation techniques—such as pitch, tempo, fundamental frequency (F0), and vocal tract length normalization (VTLN)—are applied to make the characteristics of adult speech more similar to those of child speech. The results of this study show that ASR models trained on modulated adult speech achieved lower word error rates (WER) than those trained on unmodulated adult speech. Under the best augmentation settings, WER(N) decreased by up to 0.31 percentage points for Whisper and 0.41 percentage points for Wav2Vec 2.0 (W2V-K), while CER(N) decreased by 0.51 and 0.14 percentage points, respectively. In comparative experiments, applying the most effective augmentation conditions from Whisper to the Wav2Vec2-Korean (W2V-K) model also resulted in performance improvements.

Statistics

Cite this article