Speaker Recognition Using DNN Intermediate Layer Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speaker recognition methods using deep neural networks (DNNs) fail to improve accuracy for long utterances of 1 minute or more due to the loss of abundant information from averaging short-time acoustic features, leading to insufficient recognition performance compared to i-vector methods.
Innovation Solution
A speaker recognition device and method that calculates statistical processing of acoustic features over the entire utterance using a DNN, extracting speaker features through intermediate layers, and calculating similarities with pre-stored features to enhance recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If acoustic features are averaged over short time intervals for DNN input, then processing speed is improved, but information loss occurs leading to poor accuracy for long utterances
Solution Approach 1:
The utterance is divided into multiple time intervals, and acoustic features are extracted for each interval separately rather than averaging them. This segmentation preserves the temporal characteristics and abundant information in long utterances while still enabling efficient batch processing through the DNN.
Solution Approach 2:
Instead of reducing temporal dimension through averaging, the patent maintains the temporal dimension by processing acoustic features from multiple time intervals as separate inputs to the DNN. This dimensional approach preserves information while achieving processing efficiency through the neural network's parallel processing capability.
2Speed
If DNN extracts speaker features directly from acoustic features, then recognition speed is improved, but overtraining occurs on utterance content and recording environment
Solution Approach 1:
The patent extracts speaker features from intermediate layers of the DNN rather than using the final output layer. This extraction approach separates speaker-specific characteristics from utterance content and recording environment dependencies, preventing overtraining while maintaining fast recognition speed through efficient feature extraction.
3Ease of manufacture
If traditional DNN methods are used for speaker recognition, then implementation simplicity is maintained, but accuracy for long utterances is insufficient
Solution Approach 1:
The patent performs preliminary extraction of acoustic features from utterance data before inputting them to the DNN. This preliminary processing step prepares the data in an optimized format that enhances the DNN's ability to achieve high accuracy on long utterances while maintaining implementation simplicity through a clear two-stage process.
Data Source
AI summary
A speaker recognition device according to the present disclosure includes: an acoustic feature calculator that calculates, from utterance data indicating a voice of an obtained utterance, acoustic feature of the voice of the utterance; a statistic calculator that calculates an utterance data statistic from the calculated acoustic feature; a speaker feature extractor that extracts speaker feature of a speaker of the utterance data from the calculated utterance data statistic using a deep neural network (DNN); a similarity calculator that calculates a similarity between the extracted speaker feature and pre-stored speaker feature of at least one registered speaker; and a speaker recognizer that recognizes the speaker of the utterance data based on the calculated similarity.