On November 6, Meta announced in a blog post a new open-source audio compression technology EnCodec, which claims to compress file sizes 10 times smaller than MP3 format.
Meta’s Fundamental Artificial Intelligence Research (FAIR) team has reported success in the area of AI-driven audio hyper compression, building a three-part system and training it end-to-end to compress audio data to a target size that can then be decoded using neural networks.
Meta’s new technology, EnCodec, achieves a compression ratio of about 10x compared to a 64 kbps MP3 with no quality loss.
The three parts of EnCodec include
Encoder: Takes the uncompressed data and converts it into a higher dimensional and lower frame rate representation.
Quantizer: compresses this representation to a target size, and reconstructs the original signal by training the quantizer to give the desired size (or set of sizes) while keeping the most important information. This compressed representation is stored on a disk or sent over the network and is equivalent to an a.mp3 file on a computer.
Decoder: Converts the compressed signal back to a waveform as similar as possible to the original signal. The key to lossy compression is to identify changes that are imperceptible to humans since perfect reconstruction is impossible at low bit rates. For this reason, EnCodec uses a discriminator to improve the perceptual quality of the generated samples, creating a cat-and-mouse-like game in which the discriminator’s job is to distinguish between real and reconstructed samples. The compression model tries to trick the discriminator by pushing the reconstructed samples to be perceptually more similar to the original samples in order to generate samples.
Meta said the technology doesn’t yet cover video but is currently in the planning stages with the goal of improving the audio experience for video conferencing, streaming movies, and playing games with friends in VR.
