Multitask Speech Model for Speaker Verification and Spoof Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speaker verification systems are vulnerable to spoofing attacks using synthetic speech and replayed audio, leading to unauthorized access, and require significant computational resources due to the need for separate neural network models for speaker identification and synthetic speech detection.
Innovation Solution
A multitask neural network model that integrates an audio encoder with overlain speaker identification and synthetic speech detection heads, allowing concurrent performance of both tasks without additional models, using a shared encoding for speaker embedding and synthetic speech classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If separate neural network models are used for speaker identification and synthetic speech detection, then detection accuracy is improved, but computational overhead increases
Solution Approach 1:
The patent combines speaker identification and synthetic speech detection into a single unified neural network model. The model shares common audio processing layers and only diverges at the final classification heads, reducing the number of separate models from two to one while maintaining both detection capabilities
Solution Approach 2:
The unified model performs multiple functions simultaneously: it conducts both speaker identification and synthetic speech detection through shared feature extraction layers. This multi-functional approach allows a single model to replace what previously required two separate models, reducing computational overhead while preserving detection accuracy
2Reliability
If multiple separate models are deployed for security verification, then security reliability is improved, but system complexity increases
Solution Approach 1:
The patent merges speaker identification and spoofing detection into a single integrated system that processes audio inputs through shared layers. This consolidation reduces system complexity by eliminating the need to manage, deploy, and coordinate multiple separate models while maintaining security reliability through unified anti-spoofing verification
3Adaptability or versatility
If separate processing pipelines are used for speaker verification and spoofing detection, then task specialization is improved, but processing efficiency decreases
Solution Approach 1:
The unified model is segmented into shared feature extraction layers and specialized classification heads. The shared layers handle common audio processing for both tasks, while the specialized heads maintain task-specific capabilities for speaker identification and synthetic speech detection, achieving both specialization and efficiency
Solution Approach 2:
The model achieves multi-functionality by using shared audio processing layers that serve both speaker identification and spoofing detection tasks simultaneously. This approach improves processing efficiency by avoiding redundant computations while preserving task specialization through separate output heads for each function
Data Source
AI summary
A method includes obtaining a speaker identification (SID) model trained to predict speaker embeddings from utterances spoken by different speakers, the SID model includes a trained audio encoder and a trained SID head. The method also includes receiving a plurality of synthetic speech detection (SSD) training utterances that include a set of human-originated speech samples and a set of synthetic speech samples. The method also includes training, using the trained audio encoder, a SSD head on the SSD training utterances to learn to detect the presence of synthetic speech in audio encodings encoded by the trained audio encoder. The operations also include providing, for execution on a computing device, a multitask neural network model for performing both SID tasks and SSD tasks on input audio data in parallel.