GC-LoRA cuts speech recognition errors by 10.9% with minimal parameters
New adapter architecture boosts ASR for noisy, dialectal, and child speech.
Transformer-based Speech Foundation Models have become the standard for Automatic Speech Recognition (ASR), but they often struggle when the target domain has mismatched acoustics—think noisy environments, regional dialects, or children's voices. Standard fine-tuning is expensive, and popular Parameter Efficient Fine-Tuning (PEFT) methods like LoRA only adjust global attention patterns, missing the local acoustic context needed for these niche domains.
Enter GC-LoRA (Gated Convolutional LoRA), a novel adapter architecture proposed by Natarajan Balaji Shankar and colleagues at UCLA. GC-LoRA injects Conformer-style local convolutional processing directly into the attention output projections of a pretrained transformer encoder. This lightweight adapter uses gated convolutions to efficiently model local acoustic dependencies without disturbing the pre-trained global representations. In experiments across acoustically-degraded, bandlimited, dialectal, and child speech datasets, GC-LoRA achieved Word Error Rate (WER) reductions of up to 10.9% compared to baselines while adding only minimal trainable parameters. The paper has been accepted for publication at Interspeech 2026, signaling strong peer validation. This approach could make fine-tuning ASR models far more practical for specialized real-world deployments.
- GC-LoRA reduces Word Error Rate (WER) by up to 10.9% over standard LoRA on acoustically challenging datasets (noisy, bandlimited, dialectal, child speech).
- Adds only minimal trainable parameters by injecting Conformer-style gated convolutions into the attention output projections of transformer encoders.
- Accepted at Interspeech 2026; builds on LoRA but adds local acoustic modeling missing from global attention-only PEFT methods.
Why It Matters
Enables efficient domain adaptation for speech AI in noisy or niche settings without retraining large models.