The HANCE Engine is a model inference library built with audio in mind. A large set of pre-trained models ranging from speech noise supression and de-reverberation to stem separation and recovery of missing frequency content are available.
Integrating machine learning inference directly into audio callback functions has traditionally been a challenge. Built from scratch in cross-platform C++, the HANCE Engine enables low latency and lock-free operation, optimized for seamless audio processing.
Our models are trained specifically for real-time usage, achieving low latencies down to 20 milliseconds in speech enhancement applications. Furthermore, the models are designed to be small and resource-efficient, with model file sizes down to 242 KB for the smallest noise suppression model.
HANCE.mp4
The easiest way to try out the HANCE Engine is to use the HANCE Model Player and load one of the models from the Models subdirectory in this repository. This allows you to adjust the parameters of the model in real-time, so that you can optimize these parameters for your use case.
The HANCE model player is available for MacOS and Windows, and installers can be found here. The model player is an audio plugin (AudioUnit for MacOS, VST3 for Windows), that can be used in the audio editing software of your choice for testing.
See documentation here for instructions to test with python.
The models in the Models folder have semantic names. For example, speech-denoise-48kHz-32ms.hance, signifies that this is a model that denoises speech, expects an input samplerate of 48kHz, and has a latency of 32ms. For product information, take a look at this page
All models within a family, e.g. the speech-denoise family, have similar characteristics. For denoising purposes, we recommend starting with speech-denoise-48kHz-32ms-tiny.hance to see if that meets your requirements, and move up in size if needed. We also note that if you have special requirements for particular audio circumstances, we offer can build models to better suit those circumstances. Contact us if this is the case.
The HANCE Engine supports a wide range of platforms from embedded systems to browser-based processing with WebAssembly. The use of vector arithmetic through Intel IPP, Apple vDSP, or NEON intrinsics ensures maximum performance across platforms.
- Windows 32 and 64 bit (Intel / AMD)
- Linux (Intel / AMD and ARM64)
- Mac / iOS (Intel and ARM64)
Learn more and listen to examples at HANCE.ai
- Small footprint
- Light on CPU
- No GPU requirements
- Low latency
- Cross-platform
- Easy to integrate
Please see the online API documentation here for integrating with HANCE Engine: https://hance-engine.github.io/hance-api/Documentation/
