Adapting ASR for kids without losing adult speech

If you fine‑tune an adult ASR model on child‑speech data, adult‑word error rate (WER) often climbs because the model forgets how to recognize grown‑up voices. The paper “Child ASR Adaptation with Adult Retention: An Empirical Study” examined exactly this trade‑off across Arabic and English, comparing several adaptation strategies and offering concrete guidance on which technique to pick depending on whether you prioritize child gains or adult preservation.

What the paper measured

The authors ran experiments on three families of models: encoder‑decoder, encoder‑CTC, and AudioLLM‑based ASR (including Whisper). For each family they compared:

• full fine‑tuning of all weights
• low‑rank adaptation (LoRA)
• post‑hoc weight‑space merging (LERP and TIES)

Evaluation data covered Arabic native and non‑native child speech, English MyST child speech, and adult benchmarks from MGB‑2 and LibriSpeech test‑clean. Performance was reported with WER, while the adaptation‑retention balance was quantified using three metrics: Retention Index, Child Adaptation Gain, and Adaptation Recovery.
Key findings that appear in the source:

• Child adaptation is necessary, especially for non‑native Arabic and English child speech.
• Direct adaptation (full fine‑tuning or LoRA) frequently degrades adult ASR performance.
• Bilingual adaptation (Arabic + English) is more stable than language‑specific adaptation.
• Weight‑space merging often improves the adaptation‑retention trade‑off: LERP leans toward adult retention, whereas TIES recovers stronger child gains.
• For encoder‑CTC, Whisper, and other AudioLLM‑based systems, merging yields the best balance.
• For the encoder‑decoder architecture, direct bilingual fine‑tuning still yields the lowest raw WER on child speech, even though it hurts adult retention more than merging does.

Practical takeaways for your own work

When you need to adapt an adult ASR model to child speech while keeping adult performance intact, consider the following steps based on the empirical results:

1. Pick the right model family

• If you are working with an encoder‑CTC or AudioLLM model (e.g., Whisper), start with weight‑space merging.
• If you rely on an encoder‑decoder model, be prepared to accept some adult WER increase if you go with direct bilingual fine‑tuning for the best child WER.

2. Choose a merging strategy according to your priority

• Adult‑first: Use LERP (linear interpolation) when you cannot afford a noticeable drop on adult test sets such as MGB‑2 or LibriSpeech test‑clean.
• Child‑first: Apply TIES merging when you need to maximize Child Adaptation Gain and can tolerate a modest adult WER rise.

3. When computational budget is tight, try LoRA but verify

• LoRA adapts far fewer parameters, so training is quick.
• After LoRA fine‑tuning, evaluate both child WER and adult WER on the same benchmarks; if adult retention drops beyond an acceptable threshold, switch to merging or consider a hybrid approach (LoRA followed by a small merging step).

4. Validate with the same benchmarks used in the study

• Run adult evaluation on MGB‑2 and LibriSpeech test‑clean.
• Run child evaluation on Arabic native/non‑native corpora and English MyST child speech.
• Report WER alongside the three metrics (Retention Index, Child Adaptation Gain, Adaptation Recovery) to make the trade‑off explicit.

By following this roadmap you can reproduce the study’s core insight: adaptation is required for child speech, but the method you select determines how much adult performance you sacrifice. The paper’s data show that merging—especially TIES for encoder‑CTC and AudioLLM models—often gives the most favorable balance, while direct bilingual fine‑tuning remains the top choice for raw child WER on encoder‑decoder systems when you can accept the adult cost.