Speech Enhancement Model Using Multi-Source Auxiliary Information

The patent replaces the traditional short-time Fourier transform (STFT) mechanical processing method with a deep learning-based spectrum feature extraction approach. The neural network model directly processes noisy speech signals to extract spectrum features without requiring fixed window length transformations, thereby eliminating the real-time performance degradation while maintaining accurate spectrum feature extraction capability.

The patent changes the fundamental parameter of signal processing by moving from fixed window length STFT to a flexible deep learning approach. The model adapts to varying signal characteristics dynamically, allowing spectrum feature extraction without being constrained by fixed temporal windows, thus improving real-time performance while preserving measurement precision.

The patent implements self-service by enabling the neural network model to automatically learn and extract optimal spectrum features directly from noisy speech signals for the specific task of speech enhancement. Instead of relying on manual feature design, the model adapts to the specific characteristics of the enhancement task, ensuring that the extracted features are perfectly suited for improving speech quality and intelligibility.

The patent incorporates feedback mechanisms through supervised learning where the model is trained with noisy speech as input and clean speech as target. The loss function provides continuous feedback during training, allowing the model to iteratively improve its spectrum feature extraction capability specifically for speech enhancement tasks, thereby ensuring high reliability and effectiveness.

The patent applies preliminary action by collecting and storing historical sound source information (including target speaker audio, noise audio, and reverberation data) in advance in a sound source information database. This pre-collected information is then utilized during the speech enhancement process to improve performance without significantly increasing algorithmic complexity, as the heavy lifting of data collection is performed beforehand.

The patent introduces a sound source information database as an intermediary between the noisy speech input and the enhancement process. This database stores pre-collected historical information about sound sources, which acts as a mediator to provide contextual information that improves speech enhancement performance without requiring complex real-time analysis of sound source characteristics.