google/lyra
Lyra
Lyra is a voice compression framework and low-bitrate speech codec designed to transmit high-quality audio over bandwidth-constrained networks. It utilizes an adaptive bitrate audio codec to balance audio quality and network bandwidth during active sessions.
The project employs generative audio compression, using neural networks to synthesize speech signals from minimal data and reconstruct missing audio details. This allows for high-quality voice audio reconstruction from highly compressed byte streams.
The system covers bandwidth-optimized voice over IP and real-time voice communication, focusing on low-bitrate speech compression to maintain call stability. Its capabilities include dynamic audio bitrate adjustment and voice audio processing to prevent lag and signal dropout in unstable network environments.
Features
- Adaptive Bitrate Codecs - Provides a voice compression technology that adjusts the data rate during active sessions to balance quality and bandwidth.
- Audio Sample Reconstruction - Uses generative models to reconstruct high-quality speech waveforms from model samples.
- Neural Audio Compression - Converts speech signals into a low-bitrate byte stream using neural tokenizers and generative models.
- Neural Vocoders - Translates compressed latent features back into audible speech by predicting the underlying waveform sample by sample.
- Neural Audio Reconstruction - Recreates high-quality speech signals from compressed byte streams using deep learning models.
- Latent Space Compression - Maps high-dimensional speech signals into a compact vector space for low-bitrate transmission.
- Generative Waveform Synthesis - Uses neural networks to synthesize high-quality speech waveforms from low-dimensional latent representations.
- Voice Compression Frameworks - Provides a complete set of tools for encoding and decoding speech into a low-bitrate byte stream.
- Low Bitrate Audio Coding - Reduces the size of voice audio data for transmission over networks with very limited bandwidth.
- Latent Bitrate Quantization - Controls the precision of compressed latent vectors to balance audio fidelity against available network bandwidth.
- Audio Feature Extraction - Extracts essential acoustic characteristics from voice signals to maintain intelligibility while discarding redundant data.
- VoIP Bandwidth Optimizations - Manages voice over IP calls by adjusting audio bitrates to prevent lag and signal dropout.
- Real-Time Voice and Video Communication - Streams spoken audio between devices while maintaining clarity despite slow or unstable internet connections.
- Media & Communication - Speech compression codec.
