UCLA's new Speech-LLM unifies child and adult ASR with MoE
First Speech-LLM to handle both children and adults using entropy-aware routing
Current Speech Large Language Models (Speech-LLMs) excel at adult automatic speech recognition (ASR) but falter on child speech due to acoustic and linguistic differences. A new framework from UCLA researchers tackles this with an Entropy-Aware Domain-Routed Mixture-of-Experts (MoE) approach. The system uses a Classifier-based Domain Router (C-DR) that first coarsely classifies the domain (child vs. adult, environmental conditions) and then fine-tunes routing to specialized experts. It also integrates Mixture-of-Projectors (MoP) and Mixture-of-LoRAs (MoL) to capture domain-specific variations in acoustic and language modeling.
To handle uncertainty near domain boundaries—where child and adult speech overlap—the framework introduces Entropy-Aware Routing (EAR). When the router's prediction entropy is high, it dynamically activates a shared expert trained on all domains. Evaluated on public child speech corpora, the model consistently outperformed baselines on child ASR metrics while maintaining adult ASR accuracy. Accepted to Interspeech 2026, this is the first Speech-LLM designed for unified, multi-domain ASR spanning children and adults, opening doors for more inclusive voice interfaces in education, healthcare, and accessibility.
- Classifier-based Domain Router (C-DR) uses coarse-to-fine strategy for efficient routing between child and adult speech domains
- Framework combines Mixture-of-Projectors (MoP) and Mixture-of-LoRAs (MoL) to model domain-specific acoustic and language variations
- Entropy-Aware Routing (EAR) dynamically engages a shared expert when domain prediction is uncertain, improving robustness at domain boundaries
Why It Matters
Unified child-adult ASR that works across environments, enabling more inclusive voice interfaces for education and accessibility.