New technique improves speech AI's ability to recognize domain-specific terms and names

Speech recognition systems powered by large language models often struggle when encountering specialized terminology, proper names, and domain-specific phrases. To overcome this limitation, researchers have introduced LOGIC, a framework designed to inject contextual guidance directly into the decoding process.

The core innovation#

Rather than relying on prompts alone—which offer limited control over how strongly a model biases toward suggested entities—LOGIC operates in logit space, the numerical layer where language models make their token predictions. By decoupling context injection from initial input processing, the method enables fine-grained control over how much the model should favor specified terms. This architecture sidesteps hallucination risks associated with generative error-correction approaches that can introduce spurious over-corrections.

Performance across markets#

Testing with an open-source speech language model across 11 different language locales reveals consistent gains. On average, LOGIC reduced entity-level word error rates by 9% relative to baseline performance. The approach incurs minimal computational overhead—about 2.8% additional runtime—while introducing only a marginal 0.3% increase in false alarms. When layered on top of traditional prompting techniques, LOGIC provides an additional 5% relative improvement in entity recognition accuracy.

Why this matters#

Accurate recognition of named entities, brand names, and specialized vocabulary is essential for real-world speech applications serving specific industries and regions. By offering explicit, tunable control over the biasing mechanism, LOGIC provides practitioners with a cleaner path to deployment without sacrificing inference speed or introducing spurious outputs.

The research is detailed in the full paper on arXiv.