Every grade on this site comes out of a scoring script with fixed weights and published letter bands. We fill in factor scores anchored to what vendors document and charge; the script does the arithmetic; we paste its output. No number on any page is typed by hand, and no grade has ever been adjusted after the fact — for money, for access, or for feelings.
The five criteria
Each tool is scored 0–5 (half points allowed) on five criteria. Each factor score cites the documentation that justifies it, and those citations appear in every review.
| Criterion | Weight | What earns points |
|---|---|---|
| Cleanup scope | 25% | How many distinct voice problems the tool addresses per its own docs: noise, reverb, music beds, plosives and clicks, leveling, loudness. Claims must be documented features, not adjectives. |
| Control | 20% | Strength dials, per-problem modules, previews, undo. A single magic button scores low here no matter how good the guess is — because when the guess is wrong, you are stranded. |
| Workflow | 20% | Batch processing, format support, presets, APIs, turnaround, and how naturally the tool sits in a production chain. A brilliant tool you fight is a worse tool. |
| Value | 20% | Entry price, effective per-processed-hour cost, and whether the free tier is a real trial or a tease. We do the per-hour maths in each review so you don't have to. |
| Transparency | 15% | Public prices on a page that loads, docs that state limits plainly, clear data handling. Hiding prices behind JavaScript or a sales call costs points here — this term, one tool learned that the hard way. |
The arithmetic
The weighted sum of the five factor scores is scaled to 100 and mapped to a letter:
- A ≥ 90 · A− ≥ 85 · B+ ≥ 78 · B ≥ 72 · B− ≥ 68
- C+ ≥ 62 · C ≥ 55 · C− ≥ 50 · D ≥ 40 · F below that
The script lives in our repository and ships with the site, so the day we change a weight, the change is visible and every grade recomputes together. This term's board, straight from its output: Auphonic 89.5 (A−), iZotope RX 12 81.0 (B+), LALAL.AI Voice Cleaner 65.5 (C+), ElevenLabs Voice Isolator 62.5 (C+), Waves Clarity Vx 62.0 (C+), Async Magic Dust 59.5 (C).
This term's board on the letter bands
- A≥ 90held in escrow — nobody reached it
- A−≥ 85Auphonic 89.5
- B+≥ 78iZotope RX 12 81.0
- B≥ 72
- B−≥ 68
- C+≥ 62LALAL.AI Voice Cleaner 65.5ElevenLabs Voice Isolator 62.5Waves Clarity Vx 62.0
- C≥ 55Async Magic Dust 59.5
- C−≥ 50
- D≥ 40
- Fbelow 40
What we refuse to do
- No invented benchmarks. You will not find signal-to-noise charts or listening-test percentages here unless we actually produced them, and this term we did not. Grades are computed from documented facts: features a vendor ships, limits it states, prices it charges. When a judgment call enters a factor score, the review says so in plain words.
- No hand-typed numbers. If a score on a page disagrees with the script, the script is right and the page is a bug. Tell us.
- No pay-to-play. Some links are affiliate links, always marked, always routed through
/go/so you can see them coming. Grades are computed before links are attached. Our current top grade, Auphonic, pays us nothing at all. - No grading of categories we cannot mark fairly. Krisp appears on this site as our standing recommendation for live-call noise — a different category with different physics. We monetize that recommendation and therefore keep it out of the graded cohort entirely, clearly labeled.
- No stale prices presented as fresh. Every page that states a price carries the date we verified it against the vendor's live page. If the date looks old, trust the vendor's page over ours and expect a re-check.
Why control weighs so much
A fair question — most one-button tools would score a letter higher if we dropped the control criterion, and their fans might argue simplicity is the feature. Our answer, as markers: an enhancer is destructive editing by machine judgment. When the model guesses right, controls are indeed unnecessary. But audio work is defined by the other days — the breathy narrator the model mistakes for noise, the room tone that keeps a scene alive, the sibilance that gets shaved into a lisp. On those days, a strength dial is the difference between a tool and a lottery ticket. We weight the category's weakest habit precisely because we want vendors to fix it; the first one-button tool that ships real controls with its magic will collect the A this page has been holding in escrow.
How a grade changes
Three ways, all public: a vendor ships or removes documented capability (scope/control/workflow move), changes pricing (value moves — we re-verify and re-date), or changes how honestly it communicates (transparency moves in either direction; a pricing page that starts loading numbers without JavaScript would earn its point back the same week). Re-grades land in Red Ink with a note about what moved and why.
See the rubric applied: the ranked grades, or any marked paper.