Token-Level Speaker Change Detection in Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for speaker change detection (SCD) in speech recognition are limited by segment length and require complex neural networks, leading to inaccurate and inefficient detection of speaker changes, especially in scenarios with rapid speaker transitions.
Innovation Solution
A method that converts frame-level acoustic features into token-level semantic and voiceprint features, using shared weights to detect speaker changes at the token level, enabling direct speaker-based speech recognition without post-processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If segment length-based methods are used for speaker change detection, then the system is simpler to implement, but the detection accuracy deteriorates when speakers switch quickly
Solution Approach 1:
The patent segments the speech signal into frames and extracts acoustic features at the frame level, then aggregates these features to token level. This multi-level segmentation approach allows the system to maintain simplicity while improving detection accuracy by capturing speaker changes at appropriate granularities without requiring complex post-processing
Solution Approach 2:
The patent transitions from traditional segment-level analysis to token-level analysis by introducing a new dimension of granularity. By aligning speaker change detection with linguistic tokens (words/subwords) rather than fixed time segments, the system achieves better accuracy for quick speaker switches while maintaining computational efficiency
2Measurement precision
If end-to-end neural network models are used for speaker change detection, then the detection accuracy is improved, but the model complexity and labeling requirements increase
Solution Approach 1:
The patent divides the complex end-to-end detection task into separate components: frame-level acoustic feature extraction, token-level semantic feature generation, and token-level voiceprint feature generation. Each component uses simpler, specialized models that are easier to train and require less labeling, while collectively achieving high detection accuracy
Solution Approach 2:
The patent uses acoustic features that serve multiple purposes: they are used for both semantic understanding and voiceprint extraction. This multi-functionality reduces the need for separate specialized models and extensive labeling, simplifying the overall system while maintaining high detection accuracy
3Measurement precision
If late speech-to-text processing is used, then the speaker change detection can be performed, but the processing time and complexity increase
Solution Approach 1:
The patent performs acoustic feature extraction and token-level feature aggregation in advance, before final speaker change detection. This preliminary processing organizes the data in a token-level structure that facilitates faster and more accurate speaker change detection without requiring complex post-processing of the entire speech signal
Data Source
AI summary
Embodiments of the present disclosure provide a method, apparatus, electronic device, and medium for speech processing. The method comprises generating a token-level semantic feature of target speech data based on a frame-level acoustic feature of the target speech data. The method further comprises generating a token-level voiceprint feature of the target speech data based on the frame-level acoustic feature. The method further comprises determining a token in the target speech data where speaker change occurs based on the token-level semantic feature and the token-level voiceprint feature. According to embodiments of the present disclosure, speaker change in speech data is detected at the token level in conjunction with the speaker's acoustic features and speech contents, and speaker-based speech recognition results are output directly without post-processing, simplifying the speech recognition process.