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

EnCodec uses a three-part system:

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

The system employs discriminators in a cat-and-mouse game to maintain perceptual quality. While one component tries to spot differences between original and compressed audio, another works to make the compressed version indistinguishable from the original.

Key features:

• First neural network compression system for 48 kHz stereo audio
• Exceeds CD-quality sampling rate (44.1 kHz)
• Designed for voice calls over poor connections
• Potential applications in metaverse experiences
• Works without requiring increased bandwidth

While still in research phase, EnCodec represents a significant advancement in audio compression technology, potentially enabling high-quality audio delivery regardless of network conditions.

Related Articles

Previous Articles