SNAC
Multi-scale neural audio codec: 44.1 kHz audio at 2.6 kbps with variable token rates.
About
SNAC — Multi-Scale Neural Audio Codec — is a 2024 neural codec by independent researcher Hubert Siuzdak (arXiv 2410.14411) that takes a different route from most RVQ codecs: instead of refining residuals at one frame rate, it quantizes across multiple temporal scales. The general-audio model compresses 44.1 kHz audio into four token streams at 14, 29, 57 and 115 Hz for a total of 2.6 kbps; a 32 kHz variant runs at 1.9 kbps, and a dedicated speech model handles 24 kHz audio at just 984 bits per second with under 20M parameters.
Because the coarse token streams capture long-range structure cheaply, SNAC has become a popular tokenizer choice in music and speech generation work, where fewer coarse tokens mean cheaper autoregressive modeling. The code and pretrained models are on GitHub, and it shows up as a baseline in a string of later codec papers — a decent signal for a one-person research project.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper-reported (arXiv 2410.14411): 2.6 kbps at 44.1 kHz, 1.9 kbps at 32 kHz, 984 bps for 24 kHz speech — design specs rather than quality-leaderboard scores
- No standardized public benchmark figures found in sources checked
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER metrics, metrics-v1). Evidence: published paper (arXiv 2410.14411) with full architecture and bitrate specs; open code + pretrained models (hubertsiuzdak/snac); cited as a baseline in multiple later codec papers (e.g. SAC, LongCat-Audio-Codec). Strengths: open, genuinely novel multi-scale design, real community adoption. Docked for: (1) single independent maintainer — bus-factor risk; (2) no standardized quality benchmarks found. License: not confirmed in sources checked (see repo).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Paper arXiv 2410.14411 (2024); 44.1kHz @ 2.6kbps, speech 24kHz @ 984bps
- · Open code + models (hubertsiuzdak/snac); cited as baseline in later codec papers
- · Single independent maintainer