Partial Utterance Handling in ASR Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional automatic speech recognition (ASR) systems in automated telephone call centers often misidentify and fail to process partial utterances from cordless and cellular phones due to noise, leading to errors and the need for repeated attempts, as they lack the ability to make fine-grained distinctions within the 'No Match' classification.
Innovation Solution
A method and system that utilize a linguistic filtering unit to process partial utterances through phonological, morphological, syntactic, and semantic analysis, assigning scores to suggested forms, and using context-relevant matching to derive meaningful information from rejected text, allowing for improved recognition and confirmation of intended speech.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional ASR systems use simple matching process for speech recognition, then the system complexity is low, but the recognition accuracy deteriorates when dealing with partial utterances from noisy environments
Solution Approach 1:
The patent segments the speech recognition process into multiple specialized modules: acoustic model matching, language model verification, confidence scoring, and partial utterance handling. Each module processes specific aspects of the recognition task, allowing the system to handle partial utterances systematically without overwhelming complexity
Solution Approach 2:
The patent introduces intermediary components such as the confidence score calculator and partial utterance detector that mediate between the simple matching process and the final recognition result. These intermediaries process and filter information to improve accuracy without requiring complete system redesign
2Productivity
If conventional ASR systems reject partial matches, then the error rate is reduced, but the productivity deteriorates due to increased retries and agent transfers
Solution Approach 1:
The patent applies partial matching instead of requiring complete matches. The system processes partial utterances by matching available segments against the grammar, assigns confidence scores to partial matches, and uses language models to verify plausibility. This allows the system to proceed with partial information rather than rejecting all incomplete matches, thereby improving productivity while maintaining acceptable error rates through subsequent verification steps
3Adaptability or versatility
If conventional ASR systems require complete matches with pre-defined forms, then the measurement precision is high, but the adaptability deteriorates when handling noisy speech from cordless and cellular phones
Solution Approach 1:
The patent introduces dynamic confidence scoring that adapts to the quality and completeness of the input speech. The system adjusts matching thresholds and processing depth based on detected speech quality, allowing flexible handling of noisy inputs from cordless and cellular phones while maintaining strict matching criteria when speech quality is high
Solution Approach 2:
The patent changes multiple parameters including confidence score thresholds, matching strictness levels, and processing depth based on speech quality detection. When partial utterances are detected or low confidence is assigned, the system adjusts parameters to allow partial matching while still verifying through language models, thereby adapting to noisy environments without sacrificing overall accuracy
Data Source
AI summary
A method, system and computer readable device for recognizing a partial utterance in an automatic speech recognition (ASR) system where said method comprising the steps of, receiving, by a ASR recognition unit, an input signal representing a speech utterance or word and transcribing the input signal into text, interpreting, by a ASR interpreter unit, whether the text is either a positive or a negative match to a list of automated options by matching the text with a grammar or semantic database representing the list of automated options, wherein if the ASR interpreter unit results in said positive match proceeding to a next input signal and if the ASR interpreter unit results in said negative match rejecting the text as representing said partial utterance, and processing, by a linguistic filtering unit, the rejected text to derive a correct match between the rejected text and the grammar or semantic database. And, then using the derived word for responding to the user in the next dialog turn in order to reduce or eliminate churn in the human-computer spoken dialog interaction.