A new approach for training and testing audio machine learning models
Our cloud-based synthetic data generation tool enables audio teams to work faster and produce better machine learning models. Say goodbye to collecting data using measurements and instead create much larger datasets of rich and realistic audio scenes in just a few hours. The tool is built on our proprietary simulation technology which enables modeling sound sources and receivers with great precision and acoustic propagation in complex domains. Train and test speech recognition, echo cancellation, beamforming, noise suppression, de-reverberation, blind room estimation and any other audio machine learning model.
High-fidelity acoustic simulation training data for a wide range of use
Create thousands of customized rich audio scenes by configuring sound sources, receivers, environments and materials.
Accurate sound propagation
Capture real-world acoustic behavior, including details like phase, thanks to Treble’s wave-based acoustic modeling.
Automated virtual world generation
Our platform automatically generates thousands virtual worlds with deep synthetic complexity.
Real sound sources
Model human speech, loudspeakers and other sources with arbitrary directivity patterns.
Your devices and receivers
Use arbitrary head related or device related transfer functions for receiver modeling.
Include moving sound sources with time varying source conditions.
The metadata you need
Access a wide variety of labels including ones that are not available from human labelers.
Treble's innovative technology provides a unique solution for generating high-fidelity acoustic simulation training data for a wide range of use cases
Generate large amounts of custom acoustic data sets for training, testing, and validating of models. With Treble Training Data you can generate synthetic data sets with specified distributions of source and receiver placements for different rooms. These rooms can vary in size, materials, and furnishing density, providing users with a comprehensive range of simulation scenarios. This versatility enables users to train their models