CNN Speech Wake-Up Model Reducing Computation on Low-End Chips
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech wake-up technologies using keyword-spotting methods face challenges with high error wake-up rates and computation intensity, limiting their practical application, especially on low-end devices, and require adjustments for different wake-up words, making it difficult to maintain stable performance.
Innovation Solution
A method employing a convolutional neural network (CNN) based speech wake-up model that performs windowing and framing on recorded speech, extracts features, and calculates posterior probabilities to distinguish wake-up words from non-wake-up words, reducing computation and eliminating the need for separate fillers for each wake-up word.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If keyword-spotting method with DNN model is used for speech wake-up, then speech wake-up function is achieved, but computation amount exceeds the computation power of low-end chips
Solution Approach 1:
The patent segments the speech recognition task into two distinct phases: wake-up word detection (binary classification) and general speech recognition. By using a simplified CNN model dedicated solely to wake-up detection with a smaller model size and fewer parameters, the computation amount is reduced while maintaining accuracy for the specific wake-up function.
Solution Approach 2:
The patent changes the model parameters by designing a CNN architecture with specific constraints (e.g., controlling the number of filters, kernel sizes, and layers) to reduce the total parameter count. This allows the model to run on low-end chips with limited computation power while still achieving reliable wake-up detection.
2Reliability
If keyword-spotting method with fillers is used for speech wake-up, then wake-up function is achieved, but model parameters increase and design needs to be changed for different wake-up words
Solution Approach 1:
The patent creates a universal CNN-based wake-up detector that can handle multiple different wake-up words without requiring separate models or filler designs for each word. The model learns to detect multiple wake-up words simultaneously through a unified training process, eliminating the need for keyword-spotting-specific adjustments.
Solution Approach 2:
Instead of using keyword-spotting methods that require copying and adapting DNN models for each wake-up word, the patent uses a single CNN model that can be trained once to recognize multiple wake-up words. This eliminates the repetitive model copying and adjustment process required in traditional keyword-spotting approaches.
3Productivity
If keyword-spotting method is used for speech wake-up, then speech wake-up function is achieved, but error wake-up rate increases causing user disgust
Solution Approach 1:
The patent implements a feedback mechanism where the CNN model continuously processes audio input and provides wake-up detection results. The model can be trained with feedback from misclassification examples to improve its accuracy over time, reducing error wake-up rates while maintaining fast response performance.
Data Source
AI summary
Embodiments of the present disclosure provide a method and a device for waking up via a speech based on AI and a computer device. The method includes the followings. A windowing and framing operation is performed on an online recorded speech, to obtain at least one speech frame. A feature extraction is performed on the at least one speech frame, to obtain speech features. A calculation is performed on a static speech feature contained in the at least one speech frame through a speech wake-up model based on convolutional neural network, to obtain a first posteriori probability matched with a category of non-wake-up words and a second posteriori probability matched with a category of wake-up words. It is determined that a wake-up word is contained in the online recorded speech, when the second posteriori probability is greater than or equal to a preset threshold.