Named-Entity Speech Recognition with WFST Vocabulary Decoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing on-device speech recognition systems face challenges in accurately recognizing named entities due to the need for extensive data calculation and training, leading to deteriorated accuracy and prolonged processing times, especially in environments with limited computational resources.
Innovation Solution
A speech recognition device utilizing a weighted finite state transducer (WFST) model generates a vocabulary list of named entities, segmenting them into subwords, and employing end-to-end ASR and WFST models to decode and integrate recognition results, reducing data calculation and improving accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If on-device speech recognition training is performed using pattern sentences including named entities, then speech recognition accuracy should improve, but the amount of data calculation increases excessively and processing time becomes too long
Solution Approach 1:
The patent segments the speech recognition task into two parts: (1) training a general ASR model on-device without named entities, and (2) separately managing named entity vocabulary. This segmentation allows the on-device model to be trained quickly while named entities are handled through a separate mechanism, resolving the contradiction between accuracy and training time.
Solution Approach 2:
The patent performs preliminary preparation by collecting and organizing named entity vocabulary lists before the speech recognition process. This preliminary action allows the system to have named entities ready for recognition without requiring extensive on-device training, thus improving accuracy while minimizing training time.
2Measurement precision
If on-device speech recognition training is performed using pattern sentences including named entities, then speech recognition accuracy should improve, but computational resource consumption increases excessively
Solution Approach 1:
The patent divides the computational workload by separating general speech recognition (handled by the ASR model) from named entity recognition (handled by vocabulary matching). This segmentation reduces the computational burden on the on-device processor while maintaining high accuracy for named entities.
Solution Approach 2:
The patent uses a pre-prepared vocabulary list as a reference copy for named entities during recognition. This copying approach allows the system to match named entities without performing complex training calculations on-device, thereby reducing energy consumption while maintaining recognition accuracy.
3Adaptability or versatility
If all pattern sentences combining named entities and command statements are learned on-device, then comprehensive speech recognition should be achieved, but the device complexity and processing requirements become unmanageable
Solution Approach 1:
The patent segments the language model into a general ASR component and a named entity vocabulary component. This segmentation allows the system to handle diverse speech inputs without requiring the on-device model to contain all possible named entities, thus maintaining versatility while controlling device complexity.
Solution Approach 2:
The patent creates a universal named entity vocabulary list that can be applied across different speech recognition contexts and commands. This universal approach allows the system to recognize various named entities without creating separate models for each scenario, thereby maintaining adaptability while simplifying the overall system structure.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Provided are a device for recognizing a speech input including a named entity from a user and an operating method thereof. The device is configured to: generate a weighted finite state transducer model by using a vocabulary list including a plurality of named entities; obtain a first string from a speech input received from a user, by using a first decoding model; obtain a second string by using a second decoding model that uses the weighted finite state transducer model, the second string including a word sequence, which corresponds to at least one named entity, and an unrecognized word sequence not identified as a named entity; and output a text corresponding to the speech input by substituting the unrecognized word sequence of the second string with a word sequence included in the first string.