Acoustic Model Reuse for Wakeword Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems require significant computing resources for full ASR and NLU processing, making them computationally expensive, and often necessitate a distributed computing environment to function efficiently.
Innovation Solution
The system re-uses a first acoustic model configured for limited ASR processing for wakeword detection, allowing multiple wakewords to be detected using different models, such as hidden Markov models, and associates these wakewords with different modes of operation, thereby eliminating the need for a second dedicated acoustic model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If full ASR and NLU processing is used for speech recognition, then accurate command recognition is achieved, but computational resources are significantly consumed
Solution Approach 1:
The speech processing system is segmented into two distinct paths: a lightweight wakeword detection path using acoustic models for simple keyword recognition, and a full ASR/NLU path for complex command processing. This segmentation allows the system to use minimal computational resources for routine wake-up detection while reserving full processing power for actual command execution.
Solution Approach 2:
Instead of applying full ASR and NLU processing to all audio inputs, the system applies partial processing only to audio segments that contain wakewords. The majority of audio inputs that do not trigger wakewords are discarded without expensive processing, significantly reducing overall computational resource consumption.
2Measurement precision
If a dedicated second acoustic model is created for wakeword detection, then wakeword detection accuracy is improved, but device complexity increases
Solution Approach 1:
The existing acoustic models designed for full ASR processing are made multi-functional by reusing them for wakeword detection. The same acoustic models that process complete speech commands are also employed to detect wakewords, eliminating the need for separate dedicated wakeword detection models and reducing overall system complexity.
Data Source
AI summary
A system and method performs wakeword detection and automatic speech recognition using the same acoustic model. A mapping engine maps phones/senones output by the acoustic model to phones/senones corresponding to the wakeword. A hidden Markov model (HMM) may determine that the wakeword is present in audio data; the HMM may have multiple paths for multiple wakewords or may have multiple models. Once the wakeword is detected, ASR is performed using the acoustic model.