Multi-Dimension Attention Network for Speech Spoofing Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing attention mechanisms are not applicable across multiple feature maps, limiting their effectiveness in speech processing tasks such as speech recognition and spoofing detection.
Innovation Solution
A multi-dimension attentive neural network evaluation unit is introduced, capable of receiving multi-dimension features containing two or more two-dimension feature maps, producing attention weights for each element, and generating low-dimension features or posterior probabilities for designated classes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional feature fusion or score fusion methods are used to combine multiple feature maps, then the system can utilize information from multiple feature maps, but the system cannot automatically select the most important features across different feature maps
Solution Approach 1:
The patent applies preliminary action by pre-training a separate neural network to learn attention weights for each feature map before using them in the main classification task. This pre-trained attention mechanism is then applied to automatically select important features across multiple feature maps during inference, resolving the contradiction by enabling adaptive feature selection without adding complexity to the main processing pipeline
Solution Approach 2:
The patent introduces an intermediary attention mechanism that acts as a mediator between multiple feature maps and the final classification task. This attention mechanism learns to weigh the importance of different feature maps and their elements, enabling the system to automatically select important features while maintaining a relatively simple overall architecture
2Measurement precision
If attention mechanisms are applied to enhance feature maps, then the discriminative power of features is improved, but the existing attention mechanism is only suitable for single feature map input and cannot be applied across multiple feature maps
Solution Approach 1:
The patent applies universality by designing an attention mechanism that can process multiple feature maps of different types (e.g., FFT spectrograms with different window lengths, CQT spectrograms) in a unified manner. The attention network learns to adaptively weight features across all input feature maps, making the attention mechanism universally applicable to multi-feature-map scenarios while maintaining its feature enhancement capability
Solution Approach 2:
The patent extends the attention mechanism from single feature map operation to multi-feature-map operation by adding a new dimension of processing. Instead of applying attention within a single feature map, the mechanism operates across the batch dimension to learn attention weights that apply to corresponding elements across multiple feature maps, thereby achieving multi-feature-map applicability while preserving discriminative power
3Reliability
If multiple feature maps are processed individually through separate systems, then each system can be optimized for its specific feature map, but the system cannot evaluate and select important features across different feature maps
Solution Approach 1:
The patent applies merging by combining multiple individually processed feature maps into a unified processing framework. The attention mechanism processes all feature maps simultaneously and learns to evaluate their relative importance, enabling cross-feature-map feature selection while maintaining the individual characteristics of each feature map through the shared attention weights
Data Source
AI summary
A spoofing detection apparatus 100 includes a multi-channel spectrogram creation unit 10 and an evaluation unit 40. The multi-channel spectrogram creation unit 10 extracts different type of spectrograms from speech data and integrates the different type of spectrograms to create a multi-channel spectrogram. The evaluation unit 40 evaluates the created multi-channel spectrogram by applying the created multi-channel spectrogram to a classifier constructed using labeled multi-channel spectrograms as training data and classifies it to either genuine or spoof.