Post-ASR Utterance Correction for Voice Command Misrecognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automatic speech recognition systems often misinterpret user voice utterances, leading to incorrect command execution and undermining user experience.
Innovation Solution
A multi-task ASR transformer-based large language model is trained to learn common misrecognitions of voice utterances, bypassing audio simulation and expensive decoding processes, using a sequence-to-sequence text-only model like BART-large to detect and correct recognition errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional ASR decoding processes are used to interpret voice utterances, then the system can process commands, but misrecognitions occur leading to incorrect command execution
Solution Approach 1:
The patent introduces an intermediary text correction model that sits between the ASR interpretation and command execution. This model receives the interpreted utterance and possible misrecognitions as input, then outputs a corrected interpretation. The intermediary layer filters out misrecognitions without requiring complete system redesign, thus improving reliability while maintaining manageable complexity.
Solution Approach 2:
The system creates a copy of the interpreted utterance and compares it against known misrecognitions. By generating possible misrecognitions and comparing the actual interpretation against these copies, the system can identify and correct errors without fundamentally changing the core ASR process, balancing accuracy improvement with system complexity constraints.
2Reliability
If feedback on previous interpretations is used to correct misrecognitions, then accuracy improves, but processing time and cost increase
Solution Approach 1:
The system performs preliminary action by pre-generating possible misrecognitions and storing them for quick comparison. Instead of analyzing feedback iteratively, the system prepares correction options in advance by comparing the interpreted utterance against the pre-generated misrecognition set, significantly reducing processing time while maintaining high accuracy.
Solution Approach 2:
The patent uses a cost-effective approach by implementing a lightweight text correction model that processes utterances quickly without requiring expensive, time-consuming iterative feedback loops. The system uses efficient string comparison and probability calculations rather than heavy computational resources, achieving fast correction with minimal time loss.
Data Source
AI summary
A method for detecting misrecognized voice utterances for which automated responses are presented may include inputting, to a post-speech recognition device, possible misrecognized utterances for respective utterances; inputting, to the post-speech recognition device, an interpretation of an utterance of a user, the utterance including a command; determining, using the post-speech recognition device, based on the possible misrecognized utterances, that the interpretation of the utterance is a misrecognition of the utterance; identifying, using the post-speech recognition device, based on the possible misrecognized utterances, a corrected utterance for the interpretation of the utterance, wherein the corrected utterance is different than the interpretation of the utterance; selecting, using the ASR model, the corrected utterance to replace the interpretation of the utterance; and sending the corrected utterance to a device associated with generating a response to the command to present to the user.