94M model beats 317M WavLM on child phoneme CER via age prediction auxiliary task
Joint age and phoneme training produced a 94M model that surpassed a 317M WavLM baseline and approached heavy ensembles on child speech. The resulting edge application removes cloud dependency while preserving privacy for pediatric users. Deployment record and open weights will determine whether the method generalizes beyond the DrivenData distribution.
The paper reports an auxiliary age-regression head added to a lightweight conformer backbone. Training minimized combined CTC and age MSE loss on limited child corpora. On the target test set the resulting model recorded lower character error rate than a 317M WavLM Large checkpoint fine-tuned identically. Parameter count fell from 317M to 94M, enabling on-device inference at 4–6 ms per 200 ms frame on recent mobile SoCs.
WavLM Large was pretrained on 94k hours of mixed adult speech; the age-aware objective compensates for formant and duration shifts that dominate child data. The 0.04 CER gap to competition ensembles (approximately 8.5B parameters total) is measured on the same DrivenData split released after the 2025 Phoneme Recognition Challenge. No external language model was used at inference.
Operationally the architecture ships as PhonemeTrainer, a mobile application that performs all acoustic processing locally. Latency and memory fit within Android 14 and iOS 17 constraints without cloud round-trips, satisfying COPPA and GDPR edge-processing requirements. The same training recipe transfers to low-resource adult accents where age or vocal-tract metadata is available.
Next steps include release of the 94M checkpoint and the age-augmented training script under an open license. Follow-on work will test whether the auxiliary head improves streaming ASR word error rate on the same child cohort when cascaded with a 50M-parameter decoder.
PhonemeTrainer: public 94M checkpoint reaches 10k GitHub clones within 90 days of release
Sources (3)
- [1]Primary Source(https://arxiv.org/abs/2608.10206)
- [2]WavLM Large(https://arxiv.org/abs/2110.13900)
- [3]DrivenData Phoneme Challenge(https://drivendata.org/competitions/123/)