Speech Index Pruning via Probability Thresholds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech indexing systems for spoken documents are unreliable due to errors in speech recognition, leading to inaccurate indexing and ranking of documents during searches.
Innovation Solution
The method involves indexing a speech segment by identifying at least two alternative word sequences and eliminating speech units from the index based on a comparison of their probability of appearance in the segment and a threshold value, creating a position-specific posterior lattice and subsequently building an inverted index for efficient search services.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If speech recognition is used to convert speech signals into text for indexing, then the spoken document can be indexed and searched, but errors in speech recognition lead to unreliable indexing and search results
Solution Approach 1:
The system performs preliminary actions by generating multiple alternative word sequences and their corresponding lattices before final indexing. By pre-computing position information for multiple possible word sequences and storing them in the index with associated probabilities, the system prepares multiple hypotheses in advance. During search, the system can evaluate these pre-computed alternatives to select the most reliable indexing results, thereby improving reliability while accounting for recognition errors.
2Measurement precision
If position information is stored for multiple alternative word sequences in the index, then search precision and document ranking improve, but the complexity and size of the index increases
Solution Approach 1:
The system applies local quality by storing position information and alternative word sequences selectively rather than uniformly. The index stores multiple alternative word sequences with their position information and probabilities only where necessary - specifically for words that have multiple plausible alternatives with significant probabilities. For words with high confidence single interpretations, the index stores only the most likely position information. This selective approach maintains search precision for ambiguous cases while minimizing index complexity for clear cases.
3Reliability
If multiple alternative word sequences are indexed with position information, then document ranking based on word order becomes possible, but the processing time and computational resources increase
Solution Approach 1:
The system changes parameters by introducing probability thresholds and confidence levels to control the indexing process. Instead of uniformly indexing all alternative word sequences, the system filters alternatives based on their probability scores, storing only those that meet predetermined threshold criteria. This parameter-based filtering reduces the number of alternatives that require full position information storage and processing, thereby reducing indexing processing time while maintaining accurate document ranking for the most likely word sequences.
Data Source
AI summary
A speech segment is indexed by identifying at least two alternative word sequences for the speech segment. For each word in the alternative sequences, information is placed in an entry for the word in the index. Speech units are eliminated from entries in the index based on a comparison of a probability that the word appears in the speech segment and a threshold value.