Speech Recognition Search Space Integration for Local Client Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition technologies have low accuracy in recognizing relevant local information on client devices, such as contact names, due to the general search space not considering client-specific information, leading to poor user experience.
Innovation Solution
The method generates a search space that includes preset client information by integrating it into the decoding process using a Weighted Finite State Transducer (WFST) framework, incorporating a language model, triphone state bundling list, and lexicon, and utilizes pre-trained Deep Neural Network (DNN) and Hidden Markov Model (HMM) models to calculate probabilities for accurate speech recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a general search space is used for speech recognition, then the system can operate without client-specific information, but the recognition accuracy for local client information (e.g., contact names) remains low
Solution Approach 1:
The search space is segmented into a general portion (containing common vocabulary and language rules) and a client-specific portion (containing local information like contact names). This segmentation allows the system to incorporate client-specific information to improve recognition accuracy while maintaining the operational simplicity of the general search space.
Solution Approach 2:
The patent merges the general search space with a client-specific search space into a unified search space. The client-specific search space is constructed by integrating local information (such as contact names from the client's address book) with the general language model, enabling accurate recognition of both common and client-specific terms without significantly increasing overall system complexity.
2Measurement precision
If client-specific information is integrated into the search space, then recognition accuracy for local information improves, but the complexity of generating and managing the search space increases
Solution Approach 1:
The client-specific search space is generated in advance (preliminarily) by extracting local information from the client device and integrating it with the general language model before speech recognition begins. This preliminary action reduces the complexity during actual recognition operations, as the integrated search space is already prepared and can be directly used for decoding speech signals.
3Reliability
If the search space includes preset client information, then the system can recognize local information accurately without network access, but the storage requirements and processing load on the client increase
Solution Approach 1:
The search space construction applies local quality by extracting only the specific client information relevant to the speech recognition task (such as contact names frequently used in commands) rather than storing all client data. This selective integration reduces the storage and processing requirements on the client device while maintaining the ability to accurately recognize local information offline.
Data Source
AI summary
A speech recognition method comprises: generating, based on a preset speech knowledge source, a search space comprising preset client information and for decoding a speech signal; extracting a characteristic vector sequence of a to-be-recognized speech signal; calculating a probability at which the characteristic vector corresponds to each basic unit of the search space; and executing a decoding operation in the search space by using the probability as an input to obtain a word sequence corresponding to the characteristic vector sequence.