MLow: Meta's low bitrate audio codec
Meta’s new MLow audio codec aims to deliver intelligible, relatively high-quality voice at extremely low bitrates (around 6–10 kbps), targeting users on congested, unreliable, or legacy 2G/3G and satellite links. Commenters weigh its perceived quality and robustness against established codecs like Opus, AMR-WB, Codec2, and Google’s Lyra, and debate how much benefit ultra-low bitrates bring once IP/UDP/RTP overhead and latency are factored in. Many see clear practical value for billions of users on limited data plans and weak networks, but question the lack of source code, licensing clarity, and comparison with other modern low-bitrate and ML-based codecs.
Perceived Audio Quality & Artifacts
- Many listeners find MLow clearly better than the Opus samples at very low bitrates, especially under loss.
- Others notice odd “glitchy” artifacts (e.g., voice sounding like a harp or harsh/unnatural), sometimes preferring Opus’s smoother but duller “tin can” sound at 6 kbps.
- Fascination with complex codec failure modes is noted (akin to video “datamoshing” and glitchy AI video).
Why <10 kbps Matters
- Several comments ask why ultra‑low bitrates are needed when LTE can support higher rates.
- Counterpoints: billions of users are still on 2G/3G, unstable or congested links, or extremely small data caps; lower bitrate means more talk time and more concurrent users.
- Low‑bitrate codecs are also critical for telephony backhaul, multiplexing many calls, and radio systems.
- Some are skeptical that a 10 kbps IP path is often “stable enough,” but others report real‑world use on low‑bandwidth, stable links and highlight the 30% packet‑loss demo.
Packet Overhead, Latency & VoIP Realities
- One line of argument: at typical RTP/UDP/IP overheads, header bytes dominate at these bitrates, limiting savings.
- Others respond that apps can bundle larger audio frames (e.g., >100 ms), use dynamic packet times, and voice activity detection to reduce packet rate, trading latency for bandwidth.
- Discussion of bufferbloat, fair queueing, and jitter shows that latency and loss behavior often matter more than raw bitrate.
Comparisons to Existing Codecs
- Several commenters criticize the lack of comparison with Codec2, Lyra/SoundStream, Speex, LPCNet, AMR‑WB, and classic G.729.
- Some point out that Opus was not designed to be optimal at extreme low bitrates and that Opus 1.5 with NoLACE and improved loss handling might narrow the gap.
- Robustness to bit errors versus packet loss is raised as important but not fully demonstrated; behavior under varying error models is described as unclear.
Openness, Patents & Availability
- Multiple people ask for source code, license details, and standardization plans; none are provided in the blog post.
- Concern that this becomes another IP‑encumbered codec, in contrast to patent‑free Opus; some wish Meta would donate it to an open standard to avoid patent thickets.
- As of the discussion, MLow appears to be an in‑house, production‑deployed codec with no public implementation.
Potential Applications
- Suggested uses include: WhatsApp/Messenger/Instagram calling and voice messages on poor networks, satellite and emergency voice (e.g., phone SOS), digital radio replacing AMBE, and dense backhaul links.
- Some are interested in creative uses: exploiting MLow’s artifacts for music, vocoder‑like effects, or intentional “glitch” processing.
Meta’s Motives & Reputation
- Many accept Meta’s claim that this is practical research to improve call quality and reduce data use for billions of users, especially in developing regions; internal metrics reportedly show higher engagement.
- Others remain wary of Meta: they praise its engineering and open‑source track record (LLMs, compression, frameworks) but argue this doesn’t offset broader criticisms (privacy, social harms).