Zero-Shot Voice Cloning for Synthetic Data Augmentation in Children's ASR

Abstract

Children's automatic speech recognition (ASR) lags behind adult ASR due to theacoustic mismatch between adult-trained acoustic models and children's speech, as well as the scarcity of labelled children's speech corpora. We investigate whether zero-shot text-to-speech (TTS) synthesis with voice cloning, requiring only a short reference clip per speaker, can serve as a practical substitute for real children's speech when fine-tuning large pre-trained ASR models. Using the OmniVoice multilingual TTS system we synthesise up to 16,000 child-like utterances from two English corpora (ENNI, clinical narrative samples, ages 7-12, and MyST, conversational science-tutor speech, grades 3-5) and fine-tune Whisper-small under different experimental conditions covering real data, synthetic data, cross-corpus data, and all-data combination. A systematic sensitivity analysis over speaker count, text-pool size, and utterances per speaker reveals that synthetic data alone reduces word error rate (WER) from a 19.8 % zero-shot baseline to 16.4 % on ENNI, recovering approximately 30 % of the gap to real-data fine-tuning (8.8 %). Combining synthetic with real data is essentially indistinguishable from real-data-only fine-tuning (9.2 % vs. 8.8 %). Speaker count plateaus at ~40 speakers for cross-domain transfer, while multi-speaker acoustic diversity per transcript, rather than raw text coverage, emerges as the primary determinant of cross-domain generalisation; out-of-vocabulary (OOV) rate proves to be a poor proxy for WER given Whisper's pre-trained language modelling capacity.