Continuous Speech Recognition Confidence Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Continuous speech recognition systems face challenges in accurately distinguishing between closely uttered words and managing large vocabularies without prior knowledge, leading to confusion and reduced accuracy in predicting word boundaries.
Innovation Solution
A continuous speech recognition engine employing a fine speech recognizer model, coarse sound representation generator, and coarse match generator that uses multiple resolution analysis to assign a robust confidence level to recognized words by comparing sound features to a database, allowing for improved categorization and decision-making in speech recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If continuous speech recognition is used to handle continuous voice communications, then the system can process natural flowing speech, but the system cannot accurately determine word boundaries when words are uttered closely together
Solution Approach 1:
The patent divides the continuous speech recognition problem into two separate recognition passes: a first pass that identifies potential words and a second pass that refines the recognition with confidence scoring. This segmentation allows the system to handle continuous speech while improving word boundary detection accuracy through multiple analysis stages.
Solution Approach 2:
The system performs preliminary speech recognition in a first pass to identify potential words and their time codes before conducting the second pass with confidence score calculation. This preliminary action allows the system to prepare candidate words and their temporal boundaries before final verification, improving overall accuracy.
2Adaptability or versatility
If continuous speech recognition without a priori knowledge is used, then the system can handle large vocabulary, but the system lacks basis for guessing what uttered word is trying to be identified
Solution Approach 1:
The patent implements a feedback mechanism where the first pass recognition results are used to generate confidence scores in the second pass. The system compares the acoustic match scores from the first pass with language model probabilities and temporal consistency from the second pass, providing feedback that improves word identification reliability while maintaining large vocabulary coverage.
Solution Approach 2:
The system performs preliminary analysis in the first pass to establish candidate words and their temporal boundaries before conducting the confidence scoring in the second pass. This preliminary action allows the system to gather necessary information for reliable word identification without limiting vocabulary size.
3Reliability
If paused voice command system with small vocabulary database is used, then the system has a priori knowledge for recognition, but the system cannot handle continuous flowing speech
Solution Approach 1:
The patent creates a universal speech recognition system that can handle both continuous flowing speech and paused voice commands through the same two-pass architecture. The system maintains reliability by using confidence scoring while gaining adaptability to process various speech types including continuous conversations, making it multi-functional.
Solution Approach 2:
The system dynamically adjusts its operation mode based on the input characteristics. The two-pass recognition system can adapt to both continuous speech streams and paused command inputs, changing its processing dynamics to match the input type while maintaining consistent accuracy through confidence score evaluation.
4Productivity
If single pass speech recognition is used, then the system operates quickly, but the system cannot provide robust confidence measure for each recognized word
Solution Approach 1:
The patent segments the speech recognition process into two distinct passes: a fast first pass for initial word identification and a second pass for confidence score calculation. This segmentation maintains productivity by completing the first pass quickly while improving measurement precision through the additional confidence scoring stage.
Solution Approach 2:
The system performs partial recognition in the first pass to establish candidate words, then applies excessive action in the second pass by calculating detailed confidence scores for each candidate. This approach balances productivity and precision by doing minimal necessary work in the first pass and more thorough analysis in the second pass only when needed.
Data Source
AI summary
A method, apparatus, and system are described for a continuous speech recognition engine that includes a fine speech recognizer model, a coarse sound representation generator, and a coarse match generator. The fine speech recognizer model receives a time coded sequence of sound feature frames, applies a speech recognition process to the sound feature frames and determines at least a best guess at each recognizable word that corresponds to the sound feature frames. The coarse sound representation generator generates a coarse sound representation of the recognized word. The coarse match generator determines a likelihood of the coarse sound representation actually being the recognized word based on comparing the coarse sound representation of the recognized word to a database containing the known sound of that recognized word and assigns the likelihood as a robust confidence level parameter to that recognized word.