Abbreviation Classification for OOV Speech Pronunciation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech synthesis and recognition models struggle to accurately pronounce out-of-vocabulary acronyms and initialisms, as they lack the ability to distinguish between these abbreviations and adapt to new words.
Innovation Solution
A machine learning-based approach is employed to classify input strings as either acronyms or initialisms, using trained models to generate correct pronunciations, which are then integrated into speech synthesis and recognition systems, expanding their lexicon and improving pronunciation accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional speech synthesis models are used, then speech can be generated for common vocabulary, but the models cannot accurately pronounce out-of-vocabulary acronyms and initialisms
Solution Approach 1:
The patent segments the problem of pronunciation by classifying abbreviations into distinct categories (acronyms vs. initialisms). The machine learning model divides the input space into separate classes, allowing different pronunciation rules to be applied to each segment. This segmentation enables the system to handle out-of-vocabulary words by identifying their category and applying appropriate pronunciation patterns specific to that segment.
Solution Approach 2:
The patent changes the parameter of classification granularity by introducing a binary classification system that distinguishes between acronym and initialism types. This parameter change allows the system to adapt its pronunciation generation based on the classified category, improving accuracy for out-of-vocabulary abbreviations while maintaining compatibility with conventional speech synthesis for standard vocabulary.
2Adaptability or versatility
If speech synthesis models are expanded to include more vocabulary, then coverage increases, but the models struggle to adapt to new or unknown words
Solution Approach 1:
The patent introduces a machine learning classification model as an intermediary between the input text and the speech synthesis engine. This intermediary classifies abbreviations into categories and guides the pronunciation generation process. By adding this intermediate layer, the system achieves better adaptability to new vocabulary without substantially increasing overall system complexity, as the classifier handles the adaptation logic separately from the core synthesis engine.
3Measurement precision
If automatic speech recognition models are trained on existing data, then recognition accuracy is good for known words, but they have difficulties transcribing new or unknown words
Solution Approach 1:
The patent applies preliminary classification action by using the machine learning model to categorize abbreviations before they are processed by the speech recognition system. This preliminary action of classifying the abbreviation type enables the system to prepare appropriate transcription patterns in advance, improving the ability to accurately transcribe new or unknown abbreviations while maintaining high accuracy for known vocabulary through the pre-established classification framework.
Data Source
AI summary
A method, computer system, and computer program product are provided for synthesizing and/or recognizing speech. Training data is provided to a machine learning model, wherein the training data comprises a plurality of labeled examples of acronyms and initialisms. The machine learning model is trained to classify strings into an acronym class or an initialism class. An input string is classified with the machine learning model into the acronym class or the initialism class. Based on the classifying, a pronunciation is generated for the input string.