SoundStream

lastmod 2021-08-16created 2021-08-161 edits

Features

• The first neural network codec applicable to both speech and music
  • Can run in real time on a smartphone CPU
  • Encodes a wide range of sound types, including high-quality audio and clean speech, noisy and reverberant speech, music, and ambient sounds
• Trained end-to-end with neural networks, performing compression and audio enhancement at the same time to deliver high-quality audio
  • At 3kbps, SoundStream outperforms Opus at 12 kpbs and delivers quality close to EVS at 9.6bps
  • Uses 3.2x-4x fewer bits, so transmission volume can be greatly reduced
  • Capable of excellent noise suppression
• To be integrated into Lyra, the low-bitrate speech codec released earlier this year

See also

• Google
• Multimedia Codec

Favorite site

• Google AI Blog: SoundStream: An End-to-End Neural Audio Codec