Voice Command Endpoint Detection via Tail Sampling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Voice-controlled devices in noisy environments struggle to accurately determine when a user has finished speaking a command or query due to ambient and background noise, leading to incomplete transcription and improper execution of commands.
Innovation Solution
The implementation of command boosting and tail sampling techniques, where the device processes user utterances using an ASR engine and audio cache, determines the confidence level of transcriptions and continues to capture audio after command execution to ensure complete detection of user inputs, preventing premature execution of incomplete commands.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If the device executes commands based on initial transcription of user utterance, then command execution speed is improved, but transcription accuracy deteriorates due to incomplete capture of full utterance in noisy environments
Solution Approach 1:
The system performs preliminary command detection and boosting on the first portion of the user utterance to enable early command execution. This preliminary action allows the system to potentially execute commands faster while still having the opportunity to verify completeness through subsequent tail sampling.
Solution Approach 2:
The system implements tail sampling that captures additional audio following command execution and provides feedback to verify whether the utterance was complete. This feedback mechanism allows the system to detect incomplete commands and prevent erroneous execution, thereby improving transcription accuracy without completely sacrificing execution speed.
2Measurement precision
If the device captures additional audio after command execution through tail sampling, then transcription accuracy is improved, but processing time increases
Solution Approach 1:
The system performs tail sampling for a limited duration and only processes additional audio if necessary. This partial action approach captures enough information to verify command completeness without excessively extending processing time, thus balancing accuracy improvement with time constraints.
Solution Approach 2:
The system executes commands based on initial transcription before completing full audio capture. This preliminary execution reduces perceived processing time for the user while tail sampling continues in the background to verify completeness, effectively overlapping processing steps.
3Speed
If the device uses command boosting processing rules on first portion of utterance, then command execution speed is improved, but reliability deteriorates due to premature execution of incomplete commands
Solution Approach 1:
The system uses tail sampling to provide feedback on whether the user utterance was complete before finalizing command execution. This feedback loop allows command boosting to operate on the first portion for speed while verifying reliability through subsequent audio analysis, preventing premature execution of incomplete commands.
Solution Approach 2:
The system performs preliminary command boosting on the first portion of the utterance to enable fast execution path. This preliminary action is complemented by subsequent tail sampling that verifies completeness, allowing the system to maintain high execution speed for valid commands while filtering out unreliable premature executions.
Data Source
AI summary
Provided herein are systems and methods for improved speech and command detection. For example, a user utterance may be received by a voice-enabled device. The voice-enabled device and associated system may determine that a first portion of the utterance comprises a complete command, and begin processing the command for execution. Thereafter, the device may receive an additional utterance and determine the additional utterance to be a second portion, related to the first portion, and together with the first portion comprise a different command. The device and associated system can then adjust and process the intended command.