Help & FAQ

The Refiner takes any audio file and produces a cleaner, higher-fidelity version of it. Works on AI-generated tracks from Suno or Udio, but also on podcasts, voice memos, demos, old mp3s, live recordings — anything that was compressed or is missing detail. The typical workflow is:

1. Sign in at /sign-in. The product is currently in private beta — email support@therefiner.app for access.
2. Drop a track into the dashboard. MP3 or WAV, up to 8 minutes long.
3. Pick a preset: Enhanced for fast everyday cleanup, Flow Matching for maximum detail. Optionally flip on Finalize to get the result mastered (Pro).
4. Wait for the job to finish, then download the refined WAV.

Need isolated stems instead of a refined mix? Head to the Stems page for vocals / drums / bass / other or speech / background splits. See the stem separation section below for details.

The Refiner ships two families of models. They take very different paths to the same destination, and their trade-offs are worth understanding before you spend credits.

Enhanced & Restoration — encoder–decoder with neural synthesis

These are deterministic, single-pass models. The input audio is analyzed (typically in the spectral domain), a learned network predicts what a clean version of that spectrum should look like, and a neural synthesizer renders it back to a waveform. One forward pass, no sampling, fully reproducible: feed the same file twice and you get bit-identical output.

Because the whole pipeline is one inference, it's fast — typically 10–30 seconds for a 3-minute track — and cheap (1 credit). It's excellent at correcting the systematic artifacts that lossy codecs introduce: high-frequency rolloff above ~11 kHz, quantization noise, smeared transients. The downside is that the synthesis path is itself a learned model, so very high-frequency content it produces shares some statistical fingerprints with other generative-AI audio output. If your downstream pipeline runs an AI-content detector, Restoration may still flag.

Flow Matching — iterative probabilistic reconstruction

Flow Matching is a generative model that learns a continuous trajectory from noise to clean audio, conditioned on the degraded input. Inference is iterative: the model takes many denoising steps, each one sharpening the previous estimate. Rather than "enhancing" the original signal, it reconstructs a plausible clean version of it from scratch, guided by the input.

The practical consequence is more high-frequency detail than the encoder–decoder path can produce — cymbal shimmer, breath, reverb tails, transient micro-detail. The original signal's vocoder-style artifacts get washed out in the resynthesis, which also makes Flow Matching outputs harder for AI-content detectors to flag. The cost is time and compute: every denoising step is its own forward pass, and "High" quality runs more steps than "Normal". A 3-minute track takes roughly 6–12 minutes depending on quality. 3 credits on Normal, 4 on High. Enabling AI artifact removal adds a cleaning stage of roughly the track's length first.

Finalize — automated mastering

Finalize is not a refinement model — it's a DSP mastering chain that runs after everything else. It analyzes the finished audio and applies corrective EQ, glue compression, and harmonic saturation from a two-character analog model — a valve-triode shaper for even harmonics and a transformer shaper for odd harmonics, dosed mid/side so the center gains density without narrowing the image — plus stereo polish (phase-correlation repair, a mono fold below ~150 Hz so the low end holds up on club PAs and phones, and a touch of high-frequency side air for width), then brings the track to a −9 LUFS release target under a −1 dBTP true-peak ceiling. It adds about 30 seconds to the job.

Two ways to use it: flip the Finalize (mastering) toggle on any full-track refinement preset — Enhanced, Restoration, Flow Matching, AI Artifacts (Pro and Studio, no extra credits) — or pick the standalone Finalizepreset to master a track that's already mixed (1 credit). On those refinement jobs the mastered version is the default download and the pre-master stays available as a second file; standalone delivers the mastered file, and your upload is the pre-master. It's deliberately conservative: it won't replace a mastering engineer, but it closes most of the gap between a raw AI export and a released track.

Output format

All presets output WAV. Sample rate is 44.1 kHz by default (48 kHz on Pro and Studio — better for video work and some mastering pipelines). Bit depth is 24-bit, with a 32-bit float option on Pro and Studio for Flow Matching only — useful if you're feeding the output back into a DAW for further mixing, since the extra headroom prevents intersample peaks from clipping during downstream gain stages.

Which to pick

Default to Enhanced for most material — it's fast, cheap, and good enough that the difference often isn't audible on consumer playback. Reach for Flow Matching when you need maximum fidelity (a final master, a release track, anything going through critical listening) or when you specifically want to wash out generative-AI fingerprints in the source. Whatever you pick, Finalize can master the result — that's the difference between a clean file and a release-loud one.

The Refiner can also split a track into isolated stems — useful for remixing, dropping the instrumental into a DAW, building a karaoke version, isolating a vocal to use elsewhere, or pulling speech out of a podcast for transcription. Stem separation is a Pro feature; Free users get one lifetime trial split.

Open the Stems page from the dashboard, drop a track, and pick one of three engines. Output is one WAV per stem plus a single ZIP containing all of them — handy for one-click download.

Demucs — fast 4-stem classic

The original go-to source-separation model. Splits the track into vocals, drums, bass, and "other" (everything else). Fast, well-tested, and the right default for most music.

Melband — newer transformer

Mel-Band RoFormer is a more recent architecture that produces cleaner separations — less bleed between stems and tighter isolation on vocals and percussion in particular. The trade-off is runtime: it takes longer than Demucs. Reach for it when stem quality matters more than speed (final remix, sampling, lifting a vocal you actually plan to use prominently).

Voice Cleanup — speech vs. background

A 2-stem split optimized for spoken-word audio: speech on one stem, everything else (music, room noise, applause) on the other. Useful for podcast editing, lifting an interview voice out of a noisy recording, or preparing speech-only audio for transcription.

Cost and limits

Each split costs 3 credits regardless of which engine you pick. Track length limit is the same as refinement (8 minutes). After the split completes you can preview every stem in the in-app multitrack player before downloading, and the convenience ZIP stays available alongside the individual WAVs.

Refinements are paid for with credits. The free tier gives you a small number of credits per month so you can try the service. Paid subscriptions (Hobby, Pro, Studio) and one-time credit packs add more.

Monthly subscription credits reset at the end of each billing cycle and do not roll over. Credit-pack credits do not expire and stack on top of subscription credits.

If a refinement comes out worse than the source, email support@therefiner.app with the job ID (visible in the dashboard) and we will refund the credits.

Pricing details and plan comparisons are on the pricing page.

Manage your plan, payment method, and credit balance from Account → Billing. Cancel any time — your subscription remains active until the end of the current billing period.

Source files and refined outputs are stored in your private workspace. You can delete any track from the dashboard at any time. We never use your audio to train models.

Upload fails: check the file is in a supported format (MP3 or WAV) and under 8 minutes long. Tracks longer than 8 minutes are rejected at upload — split them and refine in chunks.

Job stuck: if a Flow Matching job runs longer than 30 minutes, refresh the dashboard. If it's still pending after the refresh, email support with the job ID and we'll re-queue or refund.

Result sounds wrong: Enhanced occasionally struggles on extreme inputs (very short clips, unusual instruments). Try Flow Matching at High quality. If it's still off, email support — we read every report.

Email support@therefiner.app with the job ID (if relevant) and a short description. We typically respond within one business day.