Lexicon Generation for Speech Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems face inaccuracies due to deficient lexicons that fail to include Out Of Vocabulary (OOV) words and contain inaccurate transcriptions, particularly for proper names and pronunciations with multiple variations, leading to frustration and potential loss of goodwill in automated customer service systems.
Innovation Solution
A method for generating improved lexicons through receiving symbolic input, overgenerating potential pronunciations, identifying the best variants, and storing them in a lexicon using conversion rules and iterative retraining, which can handle multiple examples of spoken words and attributes like voice quality and accent, and models these variants in weighted networks or phoneme lists.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a traditional lexicon is used for speech recognition, then the system is simple to implement, but it fails to recognize OOV words and produces inaccurate transcriptions for proper names and words with multiple pronunciations
Solution Approach 1:
The system performs preliminary actions by overgenerating potential pronunciations for words before actual speech recognition occurs. This includes creating multiple pronunciation variants for each word in the lexicon, including OOV words, so that when speech recognition is performed, the correct pronunciation is already available in the lexicon without needing complex real-time processing
Solution Approach 2:
The system uses feedback from speech recognition results to iteratively improve the lexicon. By analyzing recognition accuracy and identifying misrecognized words, the system refines the lexicon to include better pronunciation models, particularly for OOV words and proper names, thereby improving overall recognition accuracy in subsequent iterations
2Measurement precision
If multiple pronunciation variants are stored for each word to handle variations, then recognition accuracy improves, but the lexicon size and processing complexity increase
Solution Approach 1:
The system applies local quality by generating multiple pronunciation variants selectively for specific word types that require them, such as proper names and OOV words, rather than uniformly for all words. This targeted approach improves recognition accuracy for problematic word categories while minimizing the overall increase in lexicon size
Solution Approach 2:
The system uses partial action by generating pronunciation variants only for words that are likely to have multiple pronunciations or are OOV words, rather than creating all possible variants for every word in the lexicon. This reduces the total lexicon size while still capturing the necessary pronunciation variations for accurate recognition
3Productivity
If automated speech recognition is implemented to save costs, then operational expenses decrease, but poor recognition accuracy leads to loss of goodwill
Solution Approach 1:
The system performs preliminary lexicon generation and pronunciation modeling before deployment, ensuring that the speech recognition system has access to accurate pronunciation data for OOV words and proper names from the start. This reduces recognition errors and maintains user satisfaction while preserving cost savings from automation
Solution Approach 2:
The system implements feedback mechanisms to monitor recognition accuracy and user interactions, using this information to continuously improve the lexicon and pronunciation models. This ensures that as the system processes more speech, it becomes more accurate, maintaining user goodwill while preserving automation cost benefits
Data Source
AI summary
The present disclosure relates to systems, methods, and computer-readable media for generating a lexicon for use with speech recognition. The method includes overgenerating potential pronunciations based on symbolic input, identifying potential pronunciations in a speech recognition context, and storing the identified potential pronunciations in a lexicon. Overgenerating potential pronunciations can include establishing a set of conversion rules for short sequences of letters, converting portions of the symbolic input into a number of possible lexical pronunciation variants based on the set of conversion rules, modeling the possible lexical pronunciation variants in one of a weighted network and a list of phoneme lists, and iteratively retraining the set of conversion rules based on improved pronunciations. Symbolic input can include multiple examples of a same spoken word. Speech data can be labeled explicitly or implicitly and can include words as text and recorded audio.