Adapting a speech recognition system to children's voices is a classic specialization problem: you gain on the target, but risk losing accuracy on the rest. ASR systems trained on adults struggle with child voices and non-native speakers. The apparent solution—fine-tuning on child speech—often leads to degraded performance on adult voices. The study, conducted on Arabic and English, addresses the issue with a systematic comparison of full fine-tuning, LoRA, and post-hoc weight-space merging, testing encoder-decoder, encoder-CTC, and AudioLLM-based architectures.

The data include native and non-native Arabic child speech, the English MyST corpus, and adult benchmarks such as MGB-2 and LibriSpeech test-clean. Beyond WER, the team measures the trade-off using Retention Index, Child Adaptation Gain, and Adaptation Recovery.

The first result is almost a confirmation: adaptation is necessary, especially for non-native Arabic and English child speech. But the second point shifts attention to the method. Bilingual adaptation proves more stable than language-specific adaptation. And weight-space merging often improves the trade-off, particularly for encoder-CTC, Whisper, and AudioLLM models. Within merging, LERP tends to favor adult retention, while TIES recovers stronger child speech gains. The exception is the encoder-decoder model, where direct bilingual fine-tuning remains best in raw WER.

For teams managing ASR models in self-hosted or on-premise environments, these results have precise practical implications. Post-hoc merging allows competencies to be combined without relaunching costly training cycles. There is no need to maintain separate pipelines for adults and children, nor to start from scratch when a new domain arrives. The stability of bilingual adaptation also suggests that exposing the model to multiple linguistic varieties during adaptation acts as a form of regularization, reducing the risk of degrading adult performance.

There is a structural reading: the research shifts the center of gravity from monolithic training to the composition of already-trained models. This changes incentives for builders of vertical ASR services. Instead of selling a single specialized model, vendors can offer adaptation modules to be combined depending on context. For organizations handling children's voice data, often subject to privacy and localization constraints, the ability to adapt and merge models locally reduces cloud dependence and lowers the TCO of continuous updates. This is not a detail: children's voices are among the most sensitive data to manage, and the ability to improve recognition without sending samples to external services is an increasingly frequent requirement.

A note of caution, however: no method wins on all axes. The fact that direct bilingual fine-tuning remains best in raw WER for encoder-decoder models indicates that merging is not a magic wand. The choice between LERP and TIES depends on operational goals: if the priority is not breaking service for adults, LERP is more prudent; if immediate gains on children are needed, TIES can recover more. In any case, the availability of code and models makes the comparison reproducible and lowers the barrier for those who want to test these approaches on their own data.