← Back to all sparks
O

opus

COMMS
Velocity0.0

Open interactive audio codec

Opus put neural networks inside the codec, then pushed the sample rate to 96 kHz.

audio-codecmachine-learningpacket-losshigh-resolutionstandards
Current state
Opus is at 1.6.1, a small fix release on top of 1.6. The two releases that matter in this window are 1.5, which made the first extended use of machine learning in the encoder and decoder through Deep Redundancy and Deep PLC, and 1.6, which built on that with a wideband-to-fullband bandwidth extension module, 96 kHz support under the Opus HD name, and a 24-bit API. Between and around them sit build fixes and checksum-only patches.
Where it's heading
Opus is moving in two directions that would once have looked contradictory: reconstructing audio that was never transmitted, via ML-based redundancy and loss concealment at very low bitrates, and carrying far more of it, via 96 kHz and 24-bit paths. Both extend a codec whose IETF-standard core has been frozen for years, by adding layers around it rather than changing the bitstream everyone already decodes.
Prediction
With 1.6.1 confined to minor fixes and the bandwidth-extension module newly landed, the next releases are likely to tune and optimize the existing ML modules rather than add another.

Recent moves

  1. 7mo ago

    Opus 1.6.1 fixes minor issues found after 1.6

    A two-line release fixing minor issues found after 1.6. It stabilizes the bandwidth-extension and HD work rather than adding to it.

    View source ↗
  2. 7mo ago

    Opus 1.6 adds 96 kHz audio and ML bandwidth extension

    ⚡ SPARK

    Builds directly on the 1.5 machine-learning foundation: a bandwidth-extension module, 96 kHz Opus HD and a 24-bit API. It takes Opus out of the bitrate-constrained voice niche it was designed for.

    View source ↗
  3. 1y ago

    Opus 1.5.2 fixes builds and an AVX2 crash on Windows

    Build fixes plus an AVX2 misalignment that could crash on Windows. Cleanup after the 1.5 release, with the rest of the entry being checksums.

    View source ↗
  4. 1y ago

    Opus 1.5.1 fixes the meson build

    A single meson build fix carried over from 1.5, published the same day as 1.5.2. Nothing reaches the codec itself.

    View source ↗
  5. 1y ago

    Opus 1.5 brings machine learning into the encoder and decoder

    ⚡ SPARK

    The release that set up everything after it: the first extended use of machine learning inside Opus, through Deep Redundancy and Deep PLC, with intelligible speech down to 6 kb/s. Every ML feature in 1.6 extends this foundation.

    View source ↗
  6. 3y ago

    Opus 1.4 retunes in-band FEC and adds Meson support

    Pre-ML Opus: retuned in-band FEC, a new option that enables FEC without forcing SILK mode, DTX fixes and Meson build support. Classical tuning of the kind 1.5 would soon supersede.

    View source ↗