Speech Recognition Database Search Using Phonetic Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speech recognition systems are limited by small vocabularies, which restrict access to database entries, especially in large databases with unknown structures or multilingual content, leading to inaccurate search results due to unrecognized words and pronunciation variations.
Innovation Solution
A system that uses a speech recognition unit to identify both recognized and unrecognized words (pseudo-words) by analyzing phonetic properties and semantic information, then processes these to generate search queries that can include orthographic and pronunciation alternatives, allowing for the retrieval of database entries even if they are not directly recognized.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the vocabulary size of the speech recognition system is increased to recognize more database entries, then the coverage of searchable entries is improved, but the time and resources required to search the vocabulary increase significantly
Solution Approach 1:
The patent segments the large database into multiple clusters based on phonetic similarity. Each cluster contains database entries grouped by their phonetic characteristics. This segmentation allows the system to search within smaller, organized clusters rather than a single large vocabulary, reducing search time while maintaining comprehensive coverage of database entries.
Solution Approach 2:
The patent performs preliminary organization of database entries into phonetic clusters before the actual speech recognition process. By pre-grouping entries based on their phonetic properties, the system prepares the data structure in advance, enabling faster retrieval during speech recognition without requiring a complete vocabulary search each time.
2Productivity
If the vocabulary is limited to conventional words, then the vocabulary search remains fast and resource-efficient, but many database entries cannot be accessed by voice
Solution Approach 1:
The patent changes the parameter of word representation from exact lexical matches to phonetic similarity-based matching. Instead of requiring precise word matches from a limited vocabulary, the system compares the phonetic properties of spoken input with phonetic representations of database entries, enabling recognition of entries with specialized expressions, names, and non-colloquial terms while maintaining search efficiency.
3Measurement precision
If the speech recognition system uses exact word matching, then the recognition accuracy for vocabulary words is high, but it fails to recognize words not in the vocabulary or with pronunciation variations
Solution Approach 1:
The patent introduces phonetic representations as an intermediary between the spoken input and the database entries. Instead of directly matching spoken words with database text, the system converts both to phonetic representations and compares them. This intermediary approach maintains high recognition accuracy for known words while enabling recognition of words not in the vocabulary or with pronunciation variations.
4Productivity
If the database structure is known and prepared for speech recognition, then search efficiency is improved, but this approach cannot handle databases with unknown structures or multilingual content
Solution Approach 1:
The patent creates a universal phonetic-based search mechanism that can handle any database structure and language. By converting database entries to phonetic representations and using phonetic similarity matching, the system achieves a single approach that works for known and unknown database structures, single-language and multilingual content, without requiring separate preparation for each case.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Method of recognizing speech for searching a database (106), wherein a user enters a search request as a speech input (102), comprising the following steps: processing the speech input (102) in a speech recognition step (602, 703), wherein at least one hypothesis (104) is obtained comprising recognized words and pseudo-words not recognized as words, and searching the database (106) using the at least one hypothesis (104) as a search query, wherein a list of search results (107) is obtained.