Speech Recognition Compensation Module for Error Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic speech recognition (ASR) systems often produce unintended outcomes due to inaccuracies, requiring time-consuming updates and user feedback, which is inconvenient for immediate speech recognition functionality.
Innovation Solution
An electronic device with a speech recognition compensation module that stores and matches recognition-failed natural language sentences with successfully recognized ones based on similarity and time duration, allowing for prompt and accurate speech recognition without engine updates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the accuracy of the ASR engine is improved to overcome recognition inaccuracies, then the speech recognition accuracy is improved, but it takes long time to update the ASR engine and requires developer manipulation
Solution Approach 1:
The patent introduces a compensation table as an intermediary data structure that stores mappings between recognized sentences and intended sentences. This compensation table acts as a mediator between the ASR engine and the speech recognition system, allowing error correction without modifying the ASR engine itself. The compensation table enables rapid updates of recognition accuracy by simply adding or modifying entries in this intermediate layer, rather than updating the entire ASR engine.
Solution Approach 2:
The patent segments the speech recognition system into distinct components: the ASR engine, the compensation table, and the natural language understanding module. This segmentation allows the compensation table to be independently updated and maintained without affecting the ASR engine. The compensation table stores recognition results, intended sentences, and confidence scores as separate, manageable data structures that can be updated independently.
2Measurement precision
If the ASR engine is updated to reflect various pronunciation patterns, then the speech recognition accuracy is improved, but it requires developer manipulation and does not give prompt help to users
Solution Approach 1:
The system implements self-service by automatically generating compensation table entries using confidence scores from the ASR engine. When the ASR engine produces a recognition result with a confidence score below a threshold, the system automatically creates a compensation entry mapping the recognized sentence to the intended sentence, without requiring developer intervention. This automated process allows the system to adapt to various pronunciation patterns and improve accuracy in real-time based on user feedback.
Solution Approach 2:
The patent implements a feedback mechanism where the system monitors confidence scores of ASR recognition results and uses this feedback to automatically generate compensation table entries. When low confidence is detected, the system prompts the user for the intended sentence, and this user feedback is automatically incorporated into the compensation table, enabling continuous improvement of speech recognition accuracy without developer involvement.
3Measurement precision
If the system prompts the user to assess the outcome to update the speech recognition engine, then the accuracy can be improved, but it causes user inconvenience due to time-consuming manipulation
Solution Approach 1:
The patent applies partial action by only prompting users for feedback when the ASR confidence score falls below a specific threshold. Instead of requiring user feedback for every recognition attempt, the system selectively requests input only when necessary (i.e., when confidence is low). This reduces user burden and time investment while still improving accuracy for problematic cases. The compensation table is updated incrementally with only the necessary corrections.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An electronic device and method are disclosed herein. The electronic device implements the method, including: receiving a first speech, and extracting a first text from the received first speech, in response to detecting that extraction of the first text includes errors such that a request associated with the first speech is unprocessable, storing the extracted first text, receiving a second speech and extracting a second text from the received second speech, in response to detecting that the request is processable using the extracted second text, detecting whether a similarity between the first and second texts is greater than a similarity threshold, and whether the second speech is received within a predetermined time duration of receiving the first speech, and when the similarity is greater than the threshold, and the first and second speech signals are received within the time duration, storing the first text in association with the second text.