Opus 1.5: Major update to open source audio codec introduces Machine Learning enhancements

Opus, the open-source audio codec known for its versatility, has released a major update with version 1.5. This new version brings several enhancements to audio quality, some of which are based on machine learning (ML), while maintaining full compatibility with RFC 6716.

The new version of Opus introduces features that leverage machine learning to enhance the overall audio experience. While previous versions of Opus have incorporated machine learning and even deep learning, version 1.5 marks the first time that these techniques are used to process or generate the audio signals themselves.

Despite these advanced features, Opus 1.5 is designed to run smoothly on most CPUs, including those found in mobile phones. For that reason, it should be noted that all new ML-based features in Opus 1.5 are disabled by default. These features require a compile-time switch for size considerations and a run-time switch due to CPU considerations.

In addition to machine learning enhancements, Opus 1.5 also introduces improved AVX2 optimizations, additional ARM NEON optimizations, significantly improved packet loss robustness, enhancements to low-bitrate speech quality, and support for 4th and 5th order ambisonics.