模型训练方法和基于模型的音频处理方法及对应装置
By performing temporal addition and phase information determination on the left and right channel background audio samples of stereo audio, training data is constructed to train the model, solving the problem of blurred supervision signal in mono-to-stereo rendering and achieving better stereo effect generation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING ZITIAO NETWORK TECH CO LTD
- Filing Date
- 2025-09-16
- Publication Date
- 2026-07-17
AI Technical Summary
In existing technologies, the training process of mono-to-stereo rendering models is hampered by the combination of the mid-side signal and the prediction side signal, which leads to fuzzy supervision signals and makes it difficult to clearly deconstruct the stereo effect structure, thus affecting the spatial perception restoration effect of stereo audio.
By extracting the left and right channel background audio samples of stereo audio and adding them in the time domain, the phase information is determined, and a mono input signal and a stereo effect target signal are constructed to form audio training data. Based on this data, a preset model is trained, and the phase information is preserved to enhance the stereo effect.
It improves the spatial perception reproduction of stereo audio, enhances the model's ability to learn the differences between the left and right channels, reduces amplitude loss, and improves the generation quality of stereo effects.
Smart Images
Figure CN121096362B_ABST