Abstract
General-purpose Automatic Speech Recognition (ASR) systems often struggle with domain-specific jargon and diverse accents, leading to significant drops in accuracy. This paper proposes a bio-inspired "Artificial Development" (Art-Dev) framework to repair erroneous ASR outputs. By treating sentences as biological organisms, the framework employs evolutionary principles and domain-specific ontologies to "heal" incorrect words (genotypes) and "grow" them into grammatically correct sentences (phenotypes). Experimental results in financial and retail domains demonstrate substantial improvements in Word Error Rate (WER) accuracy, with one example showing an improvement from 61.53% to 92.3%.
Introduction
Even state-of-the-art ASR systems—such as Google Now, Siri, or IBM Watson—face challenges in real-world environments. While they perform well on general conversation, their accuracy often drops to between 60% and 90% when confronted with background noise, diverse accents, or specialized domains like finance or retail.
The core issue is that these engines are trained on general corpora. When a user employs a domain-specific term, the ASR often substitutes it with a phonetically similar but contextually nonsensical general word. If this "broken" text is then fed into downstream applications like automated question-answering, the entire system fails. This research seeks a method to "repair" this output specifically for the target domain before further processing occurs.
Methodology: The Art-Dev Framework
The authors propose a novel approach called Art-Dev, inspired by Evolutionary Development (Evo-Devo) biology. The repair process is conceptualized in two distinct stages:
- Evolution (Genotype Repair): The system treats the erroneous ASR sentence as a "zygote" containing "injured genes" (incorrect words). It utilizes a domain ontology—a map of relationships between field-specific concepts—to find the "fittest" replacement. For instance, if an ASR outputs "pixels" in a retail context, the system identifies a lack of relationship between "industry" and "pixels," but finds a strong relationship and phonetic similarity with "peak sales."
- Development (Phenotype Growth): Once the correct concepts (genotypes) are identified, the system "grows" them into a full sentence (phenotype). This involves linguistic repair using Part-of-Speech (POS) tagging to identify and fix grammatical inconsistencies, such as missing determiners or incorrect pronouns, ensuring the final sentence is both semantically and syntactically sound.
Experimental Results
The researchers evaluated the Art-Dev method using ASR engines like Google Now and IBM Watson across financial and retail sales domains. The primary evaluation metric was Word Error Rate (WER) Accuracy.
As shown in the figures, the Art-Dev approach yielded significant improvements in WER Accuracy across several financial domain examples:
- Example 1 (G2 to E2): Accuracy improved from 64.2% to 75.0%.
- Example 2 (G3 to E3): Accuracy improved from 62.5% to 71.8%.
- Example 3 (G4 to E4): A dramatic jump from 61.53% to 92.3%.
In a retail example, the system successfully corrected the nonsensical phrase "industry had the pixels" to the contextually accurate "industry had the peak sales" by recognizing the phonetic and ontological link between the terms.
Discussion and Conclusion
The "aha moment" of this research is the realization that phonetic similarity alone is insufficient to fix ASR errors. By combining how a word sounds with how well it fits into a domain's specific knowledge graph (ontology), the system can make much smarter corrections. This "contextual healing" ensures the entire sentence makes sense within its specific environment.
However, the system has limitations. Its success is heavily dependent on the quality and depth of the domain-specific ontology; if a concept is missing, the evolutionary repair may fail. Additionally, the computational overhead of this two-step process remains a factor for real-time applications. Future work may focus on automated ontology generation and expanding developmental rules to handle more complex linguistic structures.
Summary updated