Opus 1.5 released: Opus gets a machine learning upgrade
Opus 1.5, a major update to the open, royalty-free audio codec, introduces targeted machine learning features to dramatically improve speech quality at very low bitrates and under extreme packet loss, while avoiding “hallucinated” words by constraining models to acoustic features rather than fully generative audio. Commenters highlight impressive demos (e.g., intelligible speech with 80–90% packet loss), debate ML ethics and bias across languages and genders, and contrast Opus with Bluetooth codecs like LC3 and proprietary formats such as xHE-AAC in terms of quality, complexity, patents, and adoption incentives. There is broad agreement that using small, efficient ML components alongside traditional DSP offers a practical path to better real-time audio on constrained devices.
ML Upgrade, Hallucinations, and Design Choices
- PLC is likened to “hallucination” since it invents plausible audio, but current Opus PLC intentionally fades out around 100 ms to avoid misleading content.
- New ML components focus on regression over acoustic features to extend vowels and smooth gaps, explicitly avoiding generation of new phonemes/words.
- The vocoder is constrained by predicted spectra to limit hallucinations further.
- Several commenters praise the “hybrid” approach: small, low‑complexity DNNs assist classic DSP instead of end‑to‑end generative models.
Packet Loss Performance and Demos
- Users are impressed by intelligibility at extreme packet loss (80–90%), noting it’s still understandable though tiring.
- 9 kbps NoLACE performance is called “insane,” with comparisons to 1990s RealAudio and other legacy low‑bitrate codecs highlighting massive quality gains over decades.
Bias, Ethics, and Dataset Composition
- Multiple threads ask whether the ML models perform differently across genders, accents, and languages.
- Training reportedly uses ~205 hours of 16 kHz speech from >900 speakers across 34 languages/dialects, with roughly balanced male/female data and speech‑only focus.
- Commenters stress that bias can exist in both ML and hand‑tuned codecs, and that codec bias can have real‑world impact (e.g., in calls or interviews).
- Some see this primarily as a technical performance concern; others frame it explicitly as an ethics/accessibility issue.
Bluetooth, LC3, and Patents
- Frustration that Opus is still not widely used in Bluetooth audio; one reason cited is competing commercial interests and royalty‑bearing codecs (aptX, LDAC).
- LC3 is described as optimized for low complexity and battery life, especially for hearing aids; superficial benchmarks show LC3 decoding about 2× faster than Opus.
- Debate over whether this complexity gap justifies LC3 over Opus, and how much robustness, DSP suitability, and vendor control played into the choice.
- Opus patent pools and licensing ambiguity raise concern, though others argue all modern codecs face similar patent‑troll risks.
Terminology and Hype
- Some appreciate the use of “ML” instead of “AI,” criticizing overuse of “AI” as marketing hype.
- Others argue the distinction is not worth worrying about, or note that ML is technically a subset of AI.