Deep Learning for Signal Processing

Deep learning offers new opportunities to develop predictive models to solve a wide variety of signal processing applications. MATLAB^® supports the entire workflow—from exploration to implementation of signal processing systems built on deep networks. You can easily get started with specialized functionality for signal processing such as:

Analyzing, preprocessing, and annotating signals interactively
Extracting features and transforming signals for training deep neural networks
Building deep learning models for real-world applications, including biomedical, audio, communications, and radar
Acquiring and generating signal datasets through hardware connectivity and simulations

3:33

Deep Learning for Signals

Free White Paper
Deep Learning for Signal Processing with MATLAB

Download now

Signal Labeling and Dataset Management

With MATLAB, you can use built-in apps and domain-specific tools that can help you prepare your signal data with tasks such as labeling and managing large volumes of signal data that are too large to fit in memory.

Learn More

Signal Labeling - Example
Label Audio Using Audio Labeler - Example
Import Data Using Datastore - Example
Waveform Segmentation Using Deep Learning - Example

Time-Frequency Transforms

Time-frequency representations describe how the spectral content in a signal evolves as a function of time. You can train deep learning networks which can identify and extract patterns from the time-frequency representations. You can also choose from a variety of techniques that can generate time-frequency representations for signals, including spectrogram, mel-frequency spectrogram, Wigner-Ville, and continuous wavelet transform (or scalograms).

Learn More

Radar Classification Using Wigner-Ville and AlexNet - Example
Radar Target Classification Using Scalograms and CNN/LSTM - Example
Cocktail Party Source Separation Using Deep Learning Networks - Example
Speech Command Recognition Using Deep Learning - Example

Preprocessing and Feature Extraction

Signal preprocessing is a crucial step for enhancing overall signal quality. You can use built-in functions and apps for cleaning up signals and remove unwanted artifacts before training a deep network. You can also extract standard and domain-specific features from signals to reduce data dimensionality for training deep learning models. You can also use automatic feature extraction techniques, such as Wavelet Scattering, to obtain low-variance features from signals and train deep networks.

Learn More

ECG Signal Classification Using Long Short-Term Memory Networks - Example
Automate Feature Extraction Using Wavelet Time Scattering - Example
Classify Gender Using Long Short-Term Memory Networks - Example
Denoise Speech Using Deep Learning Networks - Example

Signal Generation and Acquisition

Deep learning models typically require large amounts of data for training and validation. In certain situations, the availability of data can be the limiting factor in adopting deep learning techniques. With MATLAB and other add-ons for signal processing applications, you can simulate synthetic data that closely matches real-world scenarios and develop models using deep learning techniques. You can interface MATLAB with external hardware to acquire real-world data for the purposes of validating your trained models via early prototypes.

Learn More

Audio Connectivity and Audio Waveform Generation - Example
Modulation Classification with Deep Learning - Example

Network Design, Training, and Deployment

Interactively design networks, speed up training using NVIDIA^® GPUs, and get to good results faster.

Design

Import pretrained models using ONNX™, then use the Deep Network Designer app to add, remove, or rearrange layers.

Training

Whether you are using one GPU, multiple GPUs, GPUs on cloud, or NVIDIA DGX, MATLAB supports multi-GPU training with one line of code.

Deployment

Deploy deep learning models anywhere. Automatically generate code to run natively on ARM^® and Intel^® MKL-DNN. Import your deep learning models and generate CUDA^® code, targeting TensorRT and CuDNN libraries