Use the Store for licensed catalog, then jump into AI Studio when the brief needs a custom melody, voice, or alternate take.
Drop the WAV. Pick the LUFS target. The render lands in your library before your coffee is cool. One credit per track. Five-pack saves $11.95.
Same chain on every track. No magic, no model-of-the-week. You can read the version pin and the weights commit in the render log.
DeepFilterNet 3 (ONNX runtime, FP16 weights, ~9 ms per chunk on a single 4-vCPU pod)
matchering 2.0 against an in-house bank of 1,400 reference tracks across eight genres
Brick-wall true-peak limiter, ceiling at -1 dBTP, ISR 4×, look-ahead 1.5 ms
Open-Unmix runs on the bus before mastering when the box is checked
Median 1.8 s for a four-minute mix at 24-bit 48 kHz. p95 sits at 2.4 s.
No PACE iLok. No subscription gate on the mastered file. The render is yours.
Honest scope: the model is good at vocal and electronic material. For orchestral, brass-heavy big-band, or sources with extreme dynamics, our mastering team takes the file by hand. The inbox is mastering@sonichaos.io.
One render at a time, a five-pack of credits, or the Studio tier that drops the per-track fee to zero.
per master
One render. Pick the LUFS target, the file format, and whether to stem-separate first.
Master one trackfive credits
Five renders. Credits don't expire. The fifth track is effectively on the house.
Buy five creditsunlimited renders
Unlimited mastering as part of the $39.99 / month Studio plan. Bring your own LUFS preset.
See the Studio tierWorked example from the test bench. Source mix peaks at -8.2 dBFS and sits at -19.4 LUFS short-term. Target render is -14 LUFS with a true-peak ceiling at -1 dBTP.
Vocal-led pop, electronic, hip-hop, R&B, and singer-songwriter sit in the sweet spot. The reference bank covers those genres densely. Drum-and-bass and synthwave land too. For orchestral, brass-heavy big-band, or anything with extreme dynamic range, our mastering team takes the file by hand — write to mastering@sonichaos.io and we route it.
matchering 2.0 spectral-matches against a per-genre reference set. We tag every mix with CLAP and pick the top-three nearest references by cosine similarity, then average their frequency response. The match weight is exposed in the user log so you can see which references won.
Yes. The dialog at delivery offers the mastered WAV, the unmastered loudness-normalised reference, and a CSV with peak / LUFS-integrated / LUFS-short-term / true-peak readings against the target.
DeepFilterNet 3 handles hum, hiss, and HVAC noise. Hard clipping inside the source is detected and surfaced — the model won't try to invent transients that aren't there. You'll see a yellow note on the render and a one-line recommendation to back the source down by 1.5–3 dB.
Source files live in a private MinIO bucket on a Hetzner box in Helsinki. Encrypted at rest with AES-256. Deleted automatically 30 days after delivery, or the moment you click Delete in the dashboard. No third-party AI vendor sees the file.
