Voiceprint Model Update via Speaker Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voiceprint feature model methods struggle to achieve high recognition accuracy while maintaining practicability, as they require lengthy training corpora and are sensitive to variations in speaking speed and emotional fluctuations.
Innovation Solution
A method that involves obtaining an original audio stream, segmenting and clustering it using a preset algorithm to isolate individual speakers' audio streams, matching these streams with an original voiceprint feature model, and using the matched streams as additional training samples to update the model, thereby improving precision and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a longer voiceprint training corpus is used, then recognition accuracy is improved, but practicability deteriorates
Solution Approach 1:
The patent segments the training corpus into multiple short audio clips instead of requiring a single long corpus. The audio stream is divided into clips of 0.5-3 seconds each, and multiple clips are combined to form the training set. This segmentation allows the system to achieve high recognition accuracy through quantity of clips rather than length of individual clips, resolving the contradiction between accuracy and practicability.
Solution Approach 2:
The patent merges multiple short audio clips into a composite training corpus. By combining numerous short clips (each 0.5-3 seconds) into a unified training set, the system achieves the effectiveness of a long training corpus while maintaining the practical advantage of using only short individual recordings. This merging approach directly resolves the contradiction by achieving long-corpus accuracy through short-corpus combination.
2Measurement precision
If multiple manual recordings are performed to improve model precision, then recognition accuracy is improved, but user experience deteriorates
Solution Approach 1:
The patent implements automatic audio stream acquisition that utilizes existing communication audio without requiring manual user recording. The system automatically captures audio streams from communication scenarios, performs speaker segmentation and clustering, and generates training samples without user intervention. This self-service approach eliminates the poor user experience of manual multiple recordings while maintaining model precision through automatic collection of sufficient training data.
Solution Approach 2:
The patent performs preliminary speaker segmentation and clustering on audio streams before model training. By pre-processing the audio to identify and separate speaker segments, the system prepares high-quality training samples in advance without requiring users to perform multiple recording sessions. This preliminary action ensures model precision is achieved through pre-processed data rather than through burdensome user recording operations.
3Adaptability or versatility
If variations in speaking speed and intonation are accommodated to improve adaptability, then model robustness is improved, but training difficulty increases
Solution Approach 1:
The patent applies dynamic time warping (DTW) algorithm to handle variations in speaking speed and temporal characteristics. DTW dynamically aligns audio clips with different speaking speeds and temporal patterns, allowing the model to learn robust features across varying speech rates without requiring complex manual preprocessing. This dynamic approach resolves the contradiction by automatically adapting to speed variations through an efficient algorithm rather than increasing overall training complexity.
Solution Approach 2:
The patent extracts multiple acoustic features including pitch, formant frequencies, and spectral characteristics that capture intonation and speaking style variations. By transforming the audio data into these parameter representations, the model learns to recognize voiceprints across different intonation patterns and speaking styles. This parameter transformation approach improves model robustness to variations while maintaining manageable training complexity through feature engineering rather than raw data processing.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present invention is applicable to the field of voice recognition technologies, and provides a method for updating a voiceprint feature model and a terminal. The method includes: obtaining an original audio stream including at least one speaker; obtaining a respective audio stream of each speaker of the at least one speaker in the original audio stream according to a preset speaker segmentation and clustering algorithm; separately matching the respective audio stream of each speaker of the at least one speaker with an original voiceprint feature model, to obtain a successfully matched audio stream; and using the successfully matched audio stream as an additional audio stream training sample for generating the original voiceprint feature model, and updating the original voiceprint feature model. In the present invention, a valid audio stream during a call is adaptively extracted and used as the additional audio stream training sample, so as to dynamically correct the original voiceprint feature model, thereby achieving a purpose of improving precision of the voiceprint feature model and recognition accuracy under a premise of relatively high practicability.