Inverse Text Normalization via Statistical Label Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automatic speech recognition systems face challenges in inverse text normalization, particularly with deterministic approaches requiring extensive maintenance and data-driven methods needing large datasets and resources, leading to errors and inefficiencies in converting spoken-form text to written-form text.
Innovation Solution
The approach casts inverse text normalization as a labeling problem, training a statistical model to predict a sequence of labels for edit operations on spoken-form text, reducing the complexity and resource requirements by limiting the number of possible labels and operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If deterministic approaches are used for inverse text normalization, then conversion accuracy can be maintained, but extensive maintenance effort and system complexity increase
Solution Approach 1:
The patent replaces deterministic rule-based systems (mechanical) with data-driven statistical models. Instead of manually maintaining complex conversion rules, the system learns patterns from training data, automatically adapting to conversion requirements while reducing manual maintenance burden and system complexity.
Solution Approach 2:
The system changes the approach from fixed deterministic parameters to probabilistic parameters learned from data. By training statistical models on labeled data, the system dynamically adjusts conversion parameters based on learned patterns rather than relying on pre-defined rules, maintaining accuracy while reducing complexity.
2Reliability
If data-driven methods are used for inverse text normalization, then conversion accuracy improves, but large datasets and computational resources are required
Solution Approach 1:
The patent applies partial action by focusing the statistical model training on specific edit operation types and token characteristics rather than attempting to model all possible conversions. This selective approach achieves sufficient accuracy while requiring less training data and computational resources than comprehensive data-driven methods.
Solution Approach 2:
The system segments the inverse text normalization problem into discrete edit operation types (insertion, deletion, substitution, replacement) and applies targeted statistical modeling to each. This segmentation allows the system to achieve accurate conversions without requiring exhaustive training data for all possible scenarios.
3Reliability
If data-driven methods are used for inverse text normalization, then conversion accuracy improves, but computational resources and processing time increase
Solution Approach 1:
The system performs partial computation by limiting statistical model evaluation to only the necessary edit operation types for each token, rather than evaluating all possible conversions. This reduces computational overhead and energy consumption while maintaining conversion accuracy through targeted statistical analysis.
Solution Approach 2:
The patent applies different processing approaches locally based on token characteristics and position. By determining feature representations specific to each token and applying appropriate edit operations only where needed, the system reduces overall computational resource requirements while maintaining accurate conversions where they matter most.
Data Source
AI summary
Techniques for inverse text normalization are provided. In some examples, speech input is received and a spoken-form text representation of the speech input is generated. The spoken-form text representation includes a token sequence. A feature representation is determined for the spoken-form text representation and a sequence of labels is determined based on the feature representation. The sequence of labels is assigned to the token sequence and specifies a plurality of edit operations to perform on the token sequence. Each edit operation of the plurality of edit operations corresponds to one of a plurality of predetermined types of edit operations. A written-form text representation of the speech input is generated by applying the plurality of edit operations to the token sequence in accordance with the sequence of labels. A task responsive to the speech input is performed using the generated written-form text representation.