Deep Neural Network Noise Cancellation for Audio Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Computing devices face challenges in effectively canceling noise from audio data, which degrades user experience in audio and video communications and affects automatic speech recognition, due to the presence of ambient, mechanical, acoustic echo, and reverberation noise.
Innovation Solution
A noise-cancellation component using a deep neural network (DNN) with recurrent neural network (RNN) layers processes audio data to extract high-level features and generate mask data, reducing noise and isolating the utterance in the audio data, employing techniques like convolution operations and complex multiplication to produce noise-reduced output.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional noise cancellation techniques are used, then device complexity is reduced, but noise cancellation effectiveness deteriorates
Solution Approach 1:
The patent replaces traditional mechanical signal processing methods with a deep neural network-based system. The DNN model learns complex noise patterns and separation strategies automatically, substituting manual feature engineering and traditional filtering algorithms with an intelligent system that adapts to different acoustic environments, thereby improving noise cancellation effectiveness while managing complexity through automated learning.
Solution Approach 2:
The system dynamically adjusts processing parameters based on the acoustic environment. The DNN model analyzes input audio characteristics and adapts its filtering strategy in real-time, changing parameters such as filter coefficients, processing intensity, and model configuration to optimize performance for different noise conditions, speech types, and environmental contexts.
2Reliability
If deep neural network processing is applied, then noise cancellation effectiveness is improved, but processing time increases
Solution Approach 1:
The system performs preliminary processing steps before applying the full DNN model. This includes initial noise estimation, feature extraction, and preprocessing that prepares the data for more efficient DNN processing. By preparing data in advance and using lighter processing steps for preliminary analysis, the system reduces the computational burden during critical real-time processing phases.
Solution Approach 2:
The system applies different levels of processing intensity based on the situation. For less critical audio segments or when noise levels are low, the system uses reduced processing (partial action) to save time. For critical segments with high noise interference, the full DNN processing is applied. This selective approach balances effectiveness with processing time constraints.
3Object-affected harmful factors
If aggressive noise reduction is applied, then noise levels are reduced, but speech quality deteriorates
Solution Approach 1:
The DNN model applies different processing strategies to different parts of the audio signal. Instead of uniformly reducing all frequencies and components, the system selectively targets noise components while preserving speech components. The model learns to identify and protect speech-related features (such as formants, pitch contours, and temporal patterns) while removing noise, thereby achieving local optimization of both noise reduction and speech quality.
Solution Approach 2:
The system incorporates feedback mechanisms where the processed output is analyzed and used to adjust subsequent processing. The DNN model continuously monitors the relationship between input noise, processed output, and expected speech quality, using this feedback to refine its filtering strategy. This closed-loop approach ensures that noise reduction does not excessively degrade speech quality by adapting to the actual impact of processing on speech characteristics.
Data Source
AI summary
A computing device may receive audio data from a microphone representing audio in an environment of the device, which may correspond to an utterance and noise. A model may be trained to process the audio data to cancel noise from the audio data. The model may include an encoder that includes one or more dense layers, one or more recurrent layers, and a decoder that includes one or more dense layers.