ASR Language Model Segmentation for Agglutinative Languages
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automatic speech recognition (ASR) systems face challenges in recognizing and interpreting compound words in highly agglutinative languages like Gujarati and Marathi, as their language models often fail to account for the large number of valid compound words, leading to incorrect interpretations or inadequate presentation of user input.
Innovation Solution
A computer-implemented training-stage technique that segments compound words into sub-terms using a segmentation resource and validates these segmentations through pronunciation and language-specific rules, training the language model on both original and segmented terms to improve recognition and presentation of agglutinative language inputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the language model includes explicit entries for every compound word, then the recognition accuracy of compound words is improved, but the memory requirements and processing resources increase significantly
Solution Approach 1:
The patent segments compound words into sub-terms (morphemes) and trains the language model on these segments rather than on complete compound words. This allows the model to recognize and reconstruct compound words from their constituent parts, reducing the vocabulary size needed while maintaining recognition accuracy for agglutinative languages.
Solution Approach 2:
The patent introduces a segmentation resource as an intermediary component that breaks down compound words into sub-terms before training. This intermediary enables the language model to handle compound words indirectly through their segments, avoiding the need to store explicit entries for every possible compound word while still achieving accurate recognition.
2Manufacturing precision
If the language model uses manual creation of labeled training corpus, then the training precision is improved, but the ease of manufacture and scalability are reduced
Solution Approach 1:
The patent implements self-service by using the segmentation resource to automatically generate training data from unlabeled corpus. The system segments compound words and validates them through pronunciation matching and language-specific rules without requiring manual annotation, enabling automatic training data preparation.
Solution Approach 2:
The patent performs preliminary segmentation and validation of compound words before training. The segmentation resource pre-processes the corpus by breaking down compound words into valid sub-terms, and the validity tests pre-filter the data to ensure quality. This preliminary action eliminates the need for manual labeling while maintaining training precision.
3Ease of operation
If the ASR system presents segmented sub-terms separately, then the processing simplicity is improved, but the presentation accuracy of compound words deteriorates
Solution Approach 1:
The patent merges the segmented sub-terms back into complete compound words during the ASR output generation phase. After recognizing individual sub-terms from audio input, the system reassembles them into their original compound word form for presentation, maintaining both processing simplicity through segmentation and presentation accuracy through merging.
Data Source
AI summary
A training-stage technique trains a language model for use in an ASR system. The technique includes: obtaining a training corpus that includes a sequence of terms; determining that an original term in the training corpus is not present in a dictionary resource; segmenting the original term into two or more sub-terms using a segmentation resource; determining that the segmentation of the original term into the two or more sub-terms is a valid segmentation, based on two or more validity tests; and training the language model based on the terms that have been identified. A computer-implemented inference-stage technique applies the language model to produce ASR output results. The inference-stage technique merges a sub-term with a preceding term if these two terms are separated by no more than a prescribed interval of time.