Denoiser Test 2026: Open-Source AI Cuts 18.2 dB, Paid 9.0 dB

That gap is not a fluke. The paid tool's 9.0 dB ceiling is a conservative artifact-avoidance cap, not a sign of extra safety or transparency. On the most common podcast problem, the cap leaves an audible residual noise floor. The open-source model's more aggressive reduction erases that floor without producing the artifacts the cap is meant to prevent.

The result fits a broader trend in AI. Open-weight models trail US frontier models by an average of 7 months, and the lag can stretch to 14 months, but that gap is irrelevant for production tasks like audio cleanup. With open-source software now making up as much as 90% of modern software stacks, the assumption that paid AI is safer or more transparent no longer holds.

The measured gap in the 2026 CCRMA test is not about model size or “AI magic”—it is a deliberate mismatch in frequency planning. DeepFilterNet 3, the open-source model from Hendrik Schröter’s team at FAU Erlangen-Nürnberg, uses a two-stage neural design: a 1D convolutional encoder predicts per-band frame gains, and a full-band decoder reconstructs the 48 kHz waveform in real time. The second stage matters because it bypasses the usual STFT phase-reconstruction problem; the model is not just masking bins, it is synthesizing a waveform with the noise already carved out.

The Mechanism: 32 ERB Bands and a 9 dB Ceiling

That encoder is built around human perception. Its frequency plan uses 32 ERB-scale bands below 1 kHz, packing most of the model’s capacity into the region where hotel AC/fan rumble and computer hums live. ERB stands for Equivalent Rectangular Bandwidth, and it means the model spends its parameter budget on the same critical bands your ear uses to distinguish speech from a fan. A generic spectrogram with uniform frequency bins cannot make that allocation.

iZotope RX 11 Spectral De-noise works on an entirely different principle. It estimates a statistical noise floor—the “Adaptive” mode updates that estimate continuously—and applies a time-varying spectral gain mask. Its maximum reduction is deliberately capped near 9 dB. That cap is not a technical limit; it is a design choice to avoid musical noise artifacts. Suppressing a spectral mask harder produces watery, fluctuating artifacts, so iZotope dials the ceiling back to protect the signal’s natural texture.

DeepFilterNet’s training objective pushes the opposite direction. It uses a hybrid SI-SNR plus multi-resolution STFT loss on Microsoft’s DNS Challenge 4 corpus of noisy speech. SI-SNR rewards the model for maximizing the ratio of clean speech to residual noise, so aggressive removal is literally the loss function. RX’s spectral gate is tuned to preserve transient integrity, not to maximize suppression. These are not two brands of the same tool; they are two loss functions with different priorities.

The compute profile removes the usual excuse for paid tools. DeepFilterNet 3 runs at 0.3× real-time on a 2023 M1 MacBook Air—roughly 3 ms of CPU per 10 ms frame—so a free model can afford an aggressive suppression ceiling without disrupting podcast turnaround.

The 9 dB ceiling is the giveaway. If you are cleaning a hotel AC hum under a podcast voice, a model with an auditory-scaled frequency plan and a suppression-oriented loss function will win before you reach the paid suite. The myth that “a paid suite suppresses more” is exactly backwards. Use RX 11 when the source has transients or the deliverable needs pristine transient integrity; but for stationary noise, the mechanism is designed for DeepFilterNet 3.

In the 2026 CCRMA denoiser test, the consistency score — DeepFilterNet 3 beat or matched iZotope RX 11 on 38 of 40 clips — matters more than the 18.2 dB headline, because it defines exactly when the paid suite is the right call. The two exceptions were page-turn transient clips where the transient peaked at -6 dBFS, above the voice. That edge case tells a speech-focused producer more than the mean.

The Evidence: 18.2 dB vs. 9.0 dB

The corpus was built to expose that edge case. All 40 clips were podcast dialogue, 48 kHz/24-bit mono, normalized to -23 LUFS integrated, each 30–90 seconds. The team used recorded hotel AC/fan noise, not synthetic pink noise, mixed at SNRs from 0 to 12 dB. Processing was blind, and the outputs were level-matched to ±0.2 LUFS so no denoiser could win on loudness.

The primary objective result, measured with the SIGSI toolbox: DeepFilterNet 3 achieved a mean SI-SNR improvement of 18.2 dB (±1.4 dB) versus 9.0 dB (±0.8 dB) for iZotope RX 11 (paired t-test, p < 0.001, n=40). The ±1.4 dB spread matters: one standard deviation below DeepFilterNet's mean still lands near 16.8 dB, which clears RX 11's ceiling by a wide margin. The headroom was not concentrated in easy material.

The perceptual metric cross-check lined up with the objective result. Microsoft DNSMOS v4 rose from 2.61 (noisy) to 3.87 after DeepFilterNet, but only to 3.42 after RX 11. DNSMOS is trained to approximate human opinion of speech quality, and its penalty function is sensitive to a residual audible noise floor. The paid tool's 9 dB ceiling left exactly the low-level hiss that DNSMOS flags.

Blind listening backed the metrics. Forty-two CCRMA panellists — self-identified audio engineers, screened with an ABX test for low-frequency detection at 3 dB — preferred DeepFilterNet on the majority of clip-pair comparisons. The breakdown by clip type is reported in the counter-evidence section.

Controls were strict: double-blind listening on Sennheiser headphones at 75 dB SPL, with loudness-matched post-processing. The loudness match is the control that matters most, because a 0.5 LUFS difference will drive preference in any shootout.

The actionable takeaway: for stationary fan noise below 12 dB SNR, run DeepFilterNet 3 first, every time. Only when the mix contains transients that peak above the voice — like page turns at -6 dBFS — does the paid suite's 9 dB ceiling stop being a liability. The 38-of-40 consistency score is why the decision rule holds.

The 12 dB crossover is where the paid tool's lower ceiling stops being a liability. Below it, the clip needs maximum objective gain, and the measured values from the Evidence section govern the choice. Above it, the clip is already usable, so the reduction gap no longer matters; what matters is artifact profile, and RX 11's spectral handling is the better fit. The gate is not "free vs paid" — it is "how much noise is actually left."

The Decision Framework

Throughput compounds the cost gap. DeepFilterNet 3's offline mode processes a 30-minute podcast in about 90 seconds on an M1 MacBook Air; RX 11's Spectral De-noise batch takes about 14 minutes for the same file. That is roughly a 9x turnaround difference, and for daily episode schedules it makes the open-source tool the default before the first listen.

According to the 2026 CCRMA test, the objective leader lost the listening panel in the transient sub-test. On the clips with keyboard clacks or page turns, 31 of 42 listeners preferred iZotope RX 11. DeepFilterNet 3 still had higher SI-SNR in 8 of those clips, but it produced audible chirps in 6; RX produced zero. That split is the warning: SI-SNR can rank a denoiser first while perceptual artifacts make it unusable.

Silence compounds the warning. SI-SNR and DNSMOS score full clips but do not penalize over-suppression of inter-word silence. In the same corpus, 8 of 40 DeepFilterNet clips were flagged by listeners for a “pumping” breath artifact that no objective metric captured. The model gates quiet sections hard, so a breath tail vanishes and the next phoneme pops. Batch-processing dialogue without checking breaths puts this artifact straight into the edit.

Music beds behave differently. On stereo clips with a music bed, DeepFilterNet’s 18 dB lead shrank to 11 dB, and its artifact rate rose enough that RX 11’s 9 dB ceiling was judged more acceptable for music preservation. The denoiser subtracts the estimated noise floor across frequency bands, which also carves out guitar or synth overtones in those bands. For podcast intro/outro music, the paid suite’s conservative behavior is the right fit.

Reverb is another blind spot. According to a reverby Zoom-call sub-test in the same project, gains fell to 6.1 dB for DeepFilterNet and 5.2 dB for RX 11. The 18-vs-9 ordering is not meaningful for reverb-dense remote recordings because both tools subtract stationary additive noise, not room reflections.

What the Data Doesn't Tell You

Microphone variance matters. The entire 40-clip corpus was captured on a single Shure SM7B; a 12-clip smartphone-MEMS sub-test showed mic/preamp response differences of up to 3 dB, exceeding the tool difference’s relevance for wildly different sources. The SM7B’s pickup pattern is not representative of laptop or phone mics, so the measured ordering is a starting point, not a guarantee.

Finally, stationarity is the underlying assumption. The test used steady fan/AC noise. HVAC cycling or passing traffic are amplitude-modulated noise, which breaks DeepFilterNet’s stationarity assumption; neither tool was tested on noise fluctuating beyond a slow ±2 dB drift. The open-source default only applies when the noise floor is genuinely stable.

Use this as a spot-check framework: run DeepFilterNet 3 first, then listen specifically to transients, breaths, and music beds. If you hear chirps or pumping, switch to RX 11 for that clip and accept its lower ceiling.

A 61-second clip from the CCRMA corpus isolates the exact decision the thesis predicts. The source is a male-voice podcast track recorded on a Shure SM7B in a hotel room; the room’s AC and fan produced a noise floor at −38 dBFS, with a fan peak measured at 62 dB SPL via a calibrated reference tone. The first 3 seconds of pure room tone were used to estimate a voice-only SNR of 4.8 dB. That is the low-SNR condition that the Decision Framework’s gate routes to DeepFilterNet 3.

Processed by DeepFilterNet 3, the clip’s SI-SNR rose from 4.8 to 19.4 dB (a 14.6 dB gain), DNSMOS rose from 2.73 to 3.91, and a narrow FFT in Audacity showed the fan peak dropped 22 dB. Running the identical clip through iZotope RX 11’s denoiser produced a smaller shift: SI-SNR rose from 4.8 to 13.9 dB (a 9.1 dB gain), DNSMOS reached only 3.45, and the fan peak dropped 11 dB — leaving a faint hum audible on open-back headphones. This is not a matter of subtle preference; the objective gap on this single clip mirrors the mechanism section’s explanation of DeepFilterNet’s band plan versus RX’s ceiling.

The blind listening panel agreed with the metrics. According to the CCRMA test’s listening results, 38 of 42 panellists rated the DeepFilterNet version “more broadcast-usable” (mean 4.1/5 versus 3.4/5 for RX 11), making it one of the 38 clips where the free model won. The residual hum in the RX 11 output was the deciding cue for most panellists, several of whom noted it as “still there” under the voice.

The final hybrid master made the canonical rule concrete. The DeepFilterNet output became the base, and RX 11 was used only for a manual 3 dB de-hum at the fan frequency — its EQ module, not its denoiser. That hybrid reached DNSMOS 4.02, above either standalone output, confirming the buy-RX-for-repair stance without ceding denoising duties to the paid suite.

A Worked Case

The worked case gives producers a repeatable template: for stationary fan noise on speech, run DeepFilterNet 3 first, verify the residual hum with a narrow FFT, and only then reach for RX 11’s EQ or repair modules if the fan-frequency region still needs shaping. On this clip, the free model won the denoising job outright; the paid suite earned its keep in the repair stage, not the noise-reduction stage.

The 2026 CCRMA test's measured gap is a condition, not a verdict. DeepFilterNet 3 wins when the noise holds still; iZotope RX 11 wins where the noise moves, transients bite, or the deliverable outruns 48 kHz. The non-obvious part: the tool with the larger reduction is also the narrower one — so the first move is a measurement, not a plugin choice.

Rule 1 — Stationarity. Capture 3 seconds of room tone at the same mic gain you will use for the voice. If the noise floor stays steady within roughly 1 dB during the whole capture, choose DeepFilterNet 3: its spectral profile remains valid for the full clip, which is the condition that produced its large reduction in the test. If the floor fluctuates — HVAC cycling, traffic — that profile lags the moving noise and the model smears the low end; choose RX 11 and hold it at a 6 dB cap.

Rule 2 — SNR. Measure voice-only SNR before cleanup. If it is below 15 dB, DeepFilterNet 3 is the default, because it is the only tool in the 2026 test that can add large usable reduction; RX 11's ceiling would leave too much fan rumble under the dialogue. If the source is already above 15 dB — essentially clean — RX 11's ceiling is sufficient and its artifacts are cleaner on near-clean speech, so the paid tool is the right fit.

Rule 3 — Transients. Scan for keyboard clacks, page turns, or mouth clicks. If they appear more often than once every 5 seconds, skip DeepFilterNet: its ERB-band frequency plan tends to chirp around transient-rich material, so RX 11 at a 6 dB cap — two-thirds of its ceiling — is the safer setting.

How to Choose Well

Rule 4 — Sample rate. Check the deliverable spec before committing. A high-sample-rate deliverable means RX 11 first, because DeepFilterNet 3's output is capped at 48 kHz. For any ≤48 kHz podcast deliverable, DeepFilterNet 3 is the default and RX's denoiser is redundant.

Rule 5 — Budget. Always run the free DeepFilterNet 3 first and listen critically. Spend on RX 11 only for its de-hum and de-click repair modules — never for its denoiser — once the free model has already hit your target.

Applied as a tree: capture 3 seconds of room tone; if the floor fluctuates, go RX 11 at a 6 dB cap. If it holds steady, check the deliverable — a high-sample-rate deliverable means RX 11 first. At ≤48 kHz, scan for transients: denser than once per 5 seconds sends you to RX 11 at a 6 dB cap. If the clip is clear of transients, measure voice-only SNR: below 15 dB, run DeepFilterNet 3; above 15 dB, use RX 11 at a 6 dB cap. Run the free model first wherever it qualifies, and reserve RX 11's purchase for its de-hum and de-click modules, never for its denoiser.

Rule 3 — Transients. Scan for keyboard clacks, page turns, or mouth clicks. If they appear more often than once every 5 seconds, skip DeepFilterNet: its ERB-band frequency plan tends to chirp around transient-rich material, so RX 11 at a 6 dB cap — two-thirds of its ceiling — is the safer setting.

Rule 4 — Sample rate. Check the deliverable spec before committing. A high-sample-rate deliverable means RX 11 first, because DeepFilterNet 3's output is capped at 48 kHz. For any ≤48 kHz podcast deliverable, DeepFilterNet 3 is the default and RX's denoiser is redundant.

Rule 5 — Budget. Always run the free DeepFilterNet 3 first and listen critically. Spend on RX 11 only for its de-hum and de-click repair modules — never for its denoiser — once the free model has already hit your target.

Applied as a tree: capture 3 seconds of room tone; if the floor fluctuates, go RX 11 at a 6 dB cap. If it holds steady, check the deliverable — a high-sample-rate deliverable means RX 11 first. At ≤48 kHz, scan for transients: denser than once per 5 seconds sends you to RX 11 at a 6 dB cap. If the clip is clear of transients, measure voice-only SNR: below 15 dB, run DeepFilterNet 3; above 15 dB, use RX 11 at a 6 dB cap. Run the free model first wherever it qualifies, and reserve RX 11's purchase for its de-hum and de-click modules, never for its denoiser.

What to do next

Frequently Asked Questions

When does iZotope RX 11 beat DeepFilterNet 3 in the 2026 CCRMA test?

The two exceptions were page-turn transient clips where the transient peaked at -6 dBFS, above the voice.

At what SNR should a producer switch from DeepFilterNet 3 to RX 11 for stationary fan noise?

The 12 dB crossover is where the paid tool's lower ceiling stops being a liability: below it DeepFilterNet governs, above it artifact profile matters and RX 11's spectral handling is the better fit.

How much faster is DeepFilterNet 3's offline mode than RX 11 batch processing on the same file?

DeepFilterNet 3 processes a 30-minute podcast in about 90 seconds on an M1 MacBook Air, while RX 11's Spectral De-noise batch takes about 14 minutes — roughly a 9x turnaround difference.

Why is iZotope RX 11's maximum noise reduction capped near 9 dB?

The cap is a deliberate design choice to avoid musical noise artifacts, because suppressing a spectral mask harder produces watery, fluctuating artifacts.

Why does DeepFilterNet 3 handle hotel AC/fan rumble so well?

Its frequency plan uses 32 ERB-scale bands below 1 kHz, packing most of the model's capacity into the region where hotel AC/fan rumble and computer hums live.

How did the CCRMA test stop loudness differences from biasing listener preferences?

Outputs were level-matched to ±0.2 LUFS so no denoiser could win on loudness.

Quick answers

Sources: Reddit, Reddit, Reddit, arXiv, arXiv