Abstract
Contextual Automatic Speech Recognition (ASR) systems often struggle with "volume sensitivity," where performance degrades as the number of biasing phrases (hotwords) increases. This paper introduces PSC-Joint, a framework that jointly models semantic correlations at list, phrase, and token levels. By employing a "divide-and-conquer" Group Competitive Purification (GCP) strategy, the model effectively filters distractors, achieving up to a 28% relative improvement in F1 scores. Experimental results on Chinese speech corpora demonstrate that PSC-Joint maintains high accuracy and computational efficiency even with large biasing lists.
Introduction
Contextual ASR allows speech recognizers to prioritize specific phrases, such as contact names, song titles, or niche technical jargon, provided as a biasing list. While cross-attention mechanisms are the industry standard for integrating this information, they suffer from significant performance degradation as the biasing list grows—a phenomenon known as volume sensitivity. Large lists introduce "distractors" (phrases phonetically similar to the target), leading to under-biasing (missing target phrases) or over-biasing (hallucinating phrases). Existing solutions often fail to maintain performance across both small and large lists.
Methods: Purified Semantic Correlation Joint Modeling
The proposed PSC-Joint model addresses volume sensitivity by recognizing that only a small fraction of a biasing list is relevant to any given audio segment.
Multi-Level Semantic Correlation
Instead of a single attention pass, PSC-Joint calculates correlations at three granularities:
- List-level: A binary classification determining if any phrase in the list is relevant to the current audio.
- Phrase-level: Identifying specific candidate phrases within the list.
- Token-level: Matching specific characters or word pieces within phrases to the ASR's internal vocabulary.
By computing the intersection of these three levels, the model "spotlights" relevant information while suppressing noise.
Group Competitive Purification (GCP)
To manage the computational cost of multi-level modeling, the authors introduced Group Competitive Purification (GCP). This "divide-and-conquer" approach splits the biasing list into groups where phrases compete based on correlation scores. Only the highest-scoring phrases proceed to the final joint modeling stage, physically removing distractors before they can cause confusion.
Experimental Setup
The model was integrated with Paraformer, a high-performance non-autoregressive ASR model, using a collaborative decoding strategy. Evaluation was conducted on three major Chinese speech corpora:
- AISHELL-1: 178 hours of standard Mandarin.
- KeSpeech: 895 hours covering Mandarin and eight sub-dialects.
- AISHELL-NER: Used to generate realistic biasing lists (Person, Location, Organization) ranging from 51 to 1,196 phrases.
Results and Discussion
PSC-Joint-P (the purified version) demonstrated significant improvements over baseline contextual models. On AISHELL-1, it achieved a 21.34% relative F1 score improvement, while on KeSpeech, the improvement reached 28.46%.
Robustness and Efficiency
A key finding is the model's stability across varying list sizes. While baseline performance typically degrades as lists grow, PSC-Joint remains robust. Furthermore, the GCP mechanism achieved a 92.06% retention rate of target phrases while ensuring that inference speed scales nearly linearly with list length, as shown in the Real-Time Factor (RTF) analysis.
Conclusion
PSC-Joint effectively mitigates the volume sensitivity of contextual ASR through multi-level semantic intersection and competitive purification. While the current study assumes a single biasing phrase per sentence and focuses on Chinese Mandarin, the architecture provides a scalable foundation for high-volume industrial ASR applications. Future work will explore multi-entity scenarios and broader language support.
Summary updated