Hugging Face Releases Child ASR Adaptation Framework: Addressing Adult Speech Forgetting in Child Speech Recognition

Summary:Hugging Face has released a new study on child Automatic Speech Recognition (ASR) adaptation, addressing the challenge of adult speech forgetting when adapting adult ASR models to child speech. The research compares full fine-tuning, LoRA, and post-hoc weight-space merging across encoder-decoder, encoder-CTC, and AudioLLM-based ASR systems. Metrics such as Retention Index, Child Adaptation Gain, and Adaptation Recovery are introduced to quantify the trade-off between adaptation and retention. Re

---

Background and Challenges

Automatic Speech Recognition (ASR) systems often underperform for children and non-native speakers, and adapting adult ASR models to child speech can lead to a phenomenon known as 'adult speech forgetting,' where the model's performance on adult speech degrades. To address this, Hugging Face's research team has proposed a new adaptation framework and conducted an in-depth experimental study.

Methodology and Experiments

The team compared the following methods:

• Full Fine-Tuning: Fine-tuning the entire model.
• Low-Rank Adaptation (LoRA): Updating parameters through low-rank matrix decomposition.
• Weight-Space Merging: Dynamically merging weights during adaptation.

Experiments used Arabic native and non-native child speech, English MyST child speech, and adult benchmarks from MGB-2 and LibriSpeech test-clean. Evaluation metrics included Word Error Rate (WER), Retention Index, Child Adaptation Gain, and Adaptation Recovery.

Key Findings

1. Necessity of Child Adaptation: Child adaptation is particularly important for non-native Arabic and English child speech, but direct adaptation often reduces adult ASR performance.
2. Advantage of Bilingual Adaptation: Bilingual adaptation is more stable than language-specific adaptation.
3. Effectiveness of Weight-Space Merging: Weight-space merging methods are effective in encoder-CTC, Whisper, and AudioLLM architectures, with LERP favoring adult retention and TIES recovering stronger child gains.
4. Encoder-Decoder Model Performance: For the encoder-decoder model, direct bilingual fine-tuning remains strongest in raw WER.

Industry Impact and Recommendations

This research provides new technical paths for the child speech recognition field, especially in multilingual and multimodal environments. Developers can consider the following recommendations:

• Prioritize Bilingual Adaptation: When resources permit, prioritize bilingual adaptation for more stable performance.
• Combine with Weight-Space Merging Methods: In encoder-CTC, Whisper, and AudioLLM architectures, combining with weight-space merging methods can effectively enhance adaptation effects.
• Focus on Adult Speech Retention: During child adaptation, focus on adult speech retention to avoid performance degradation.

Future Directions

The research team plans to further optimize adaptation methods and explore more cross-lingual and cross-modal adaptation strategies to improve ASR system performance in complex environments.

---

Source: ArXiv NLP/LLM (cs.CL) (2026-10-08)

— END —

Tags: #Hugging Face Follow #ASR Follow #Child Speech Recognition Follow #Multilingual Models Follow #AI Adaptation Follow