语音识别模型的训练方法及装置、存储介质及电子设备
By introducing real noise data into the pre-training stage of the speech recognition model, the model's ability to learn audio features in noisy environments is improved. Furthermore, by constructing a target speech recognition model through masking representation and fine-tuning, the problem of poor speech recognition performance in noisy environments is solved, achieving higher robustness and recognition accuracy.
CN116343781BActive Publication Date: 2026-07-17JD DIGITS HAIYI INFORMATION TECHNOLOGY CO LTD
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JD DIGITS HAIYI INFORMATION TECHNOLOGY CO LTD
- Filing Date
- 2023-03-23
- Publication Date
- 2026-07-17
AI Technical Summary
Technical Problem
Existing speech recognition models are not robust in noisy environments, resulting in poor recognition performance.
Method used
By introducing real noise data for pre-training, the model's ability to learn audio features in noisy environments is improved. The noise data is used as a mask representation to ensure that the encoding strategy is consistent between the pre-training and subsequent training stages. The initial speech recognition model is fine-tuned by combining labeled speech data to construct the target speech recognition model.
Benefits of technology
This improved the robustness and recognition accuracy of the speech recognition model in noisy environments, and enhanced the model's performance under noisy conditions.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN116343781B_ABST
Abstract
本公开提供了一种语音识别模型的训练方法及相关设备。该方法包括：获取第一语音数据和噪声数据，以及获取第二语音数据及其标签；通过初始语音特征提取模型处理第一语音数据和噪声数据，获得第一语音数据的原始语音特征向量，以及获得经过噪声掩码处理的掩码语音特征向量；利用原始语音特征向量和掩码语音特征向量训练初始语音特征提取模型获得目标语音特征提取模型；将目标语音特征提取模型与初始全连接层连接以构建初始语音识别模型，并通过初始语音识别模型处理第二语音数据获得第二语音数据的识别结果；根据识别结果和标签训练初始语音识别模型，获得目标语音识别模型。该方法引入了真实噪声数据，可以提升模型在噪声中学习音频特征的能力。
Need to check novelty before this filing date? Find Prior Art