Meta has introduced an AI-powered audio compression technology called 'EnCodec' that achieves 10x better compression than traditional MP3 format, while maintaining high audio quality.

EnCodec works through a three-part system:

• An encoder converts uncompressed audio into a lower frame rate representation
• A quantizer compresses this representation while preserving essential data
• A decoder reconstructs the audio in real-time using neural networks

The system uses discriminators to improve audio quality by playing a "cat-and-mouse game" - the discriminator identifies differences between original and reconstructed samples, while the compression model tries to generate samples that fool these discriminators.

Key achievements:

• First neural network application for 48 kHz stereo audio compression
• Exceeds CD quality (44.1 kHz sampling rate)
• Delivers high-quality voice calls over poor network connections
• Potential applications in metaverse experiences

While still in research phase, EnCodec demonstrates significant potential for delivering high-quality audio across networks regardless of conditions, particularly beneficial for metaverse applications and voice communications.

Related Articles

Previous Articles