Reproduce “Improving Synthetic Data Training for Contextual Biasing Models with a Keyword-Aware Cost Function” on OpenResearch

About this paper

Chin Yuen Kwok, Jia Qi Yip, Eng Siong Chng

Researchers at Nanyang Technological University introduced a keyword-aware cost function for training contextual biasing models with synthetic data, achieving a 16.6% relative Word Error Rate reduction for Whisper-small on the NSC-Part-2 dataset. The new training objective directly addresses overfitting to synthetic artifacts, improving rare word recognition for large ASR foundation models.

Key insights

• Standard ASR cross-entropy loss is ineffective for training contextual biasing modules on synthetic data when the base ASR model is already well-optimized, leading to module overfitting and no performance gain.
• Explicitly guiding the biasing module with separate loss terms for "when to bias" and "what to bias" overcomes synthetic data overfitting, resulting in substantial improvements in rare word recognition.
• The binary classification objective for `P_gen_i` enhances interpretability through False Acceptance Rate (FAR) and True Acceptance Rate (TAR) metrics, allowing for analysis and tuning of biasing aggressiveness.

Summary by alphaXiv.