# Should you de-reverb or de-noise first?

Hannah Morgan · August 22, 2026

> The Direct Answer: De-Noise First, De-Reverb Second If you only read one paragraph of this article, make it this one: in the overwhelming majority of...

## The Direct Answer: De-Noise First, De-Reverb Second

If you only read one paragraph of this article, make it this one: in the overwhelming majority of real-world cases, you should remove broadband background noise (hiss, hum, fan noise, air conditioning, traffic) before you attempt to remove reverberation or echo. The reasoning is rooted in how both algorithms work. De-noise tools build a statistical profile of steady-state noise and subtract it spectrally; they need a relatively clean signal model to do that accurately. De-reverb tools, by contrast, work by identifying late reflections and diffuse energy trails and suppressing them — a task that becomes dramatically harder when the algorithm cannot distinguish between a reverb tail and a layer of hiss sitting on top of it.

**Also worth reading:** [How to remove reverb from podcast audio effectively in 2026?](https://audobox.com/knowledge/how_to_remove_reverb_from_podcast_audio_effectively_in_2026.php) · [LALAL.AI vs iZotope RX: which audio cleanup and stem separation tool should you actually use in 2026?](https://audobox.com/knowledge/lalalai_vs_izotope_rx_which_audio_cleanup_and_stem_separation_tool_should_you_actually_use_in_2026.php) · [What are voice actor AI contract clauses and how do they work?](https://audobox.com/knowledge/what_are_voice_actor_ai_contract_clauses_and_how_do_they_work.php)

There is one important exception worth stating up front. If your recording suffers from severe, obvious room echo — the kind where words are smeared together and intelligibility is already compromised — some engineers prefer to run a gentle de-reverb pass first, because heavy noise reduction applied on top of a smeared signal can produce metallic, watery artifacts that no second pass will fix. But this applies only to extreme cases. For typical podcast recordings, voiceovers, dialogue from location shoots, and vocal takes captured in untreated rooms, the standard order is: de-noise, then de-reverb, then any corrective EQ or restoration steps.

This ordering has been validated repeatedly in professional practice. iZotope's RX documentation and training materials consistently recommend spectral de-noise before de-reverb modules, and the RX 12 shootout conducted by Nick Lear for ProVideo Coalition demonstrated that module order materially changes results when stacking Dialogue Isolate, De-noise, and De-reverb passes. Tools like Waves Clarity Vx DeReverb Pro — which won a Technical Excellence & Creativity (TEC) Award — and BandLab's Voice Cleaner all assume a reasonably noise-free input to perform at their best.

## Why Order Matters: How Each Algorithm Actually Works

Understanding why sequence matters requires a basic grasp of what each processor is doing under the hood. A de-noise plugin — whether it is a classic spectral subtractor like RX's Spectral De-noise, an AI-driven tool like Lalal.ai's background noise remover, or a learned-model system like Dialogue Isolate — estimates the spectral characteristics of the noise floor and attenuates content matching that profile. It relies on the assumption that the noise is statistically stationary: the hiss of a preamp, the 50/60 Hz hum of a bad power supply, the broadband wash of an HVAC system. These signals have stable frequency signatures over time, which makes them learnable and removable.

A de-reverb processor faces a fundamentally different problem. Reverberation is not stationary; it is a time-domain phenomenon where early reflections arrive milliseconds after the direct sound, followed by a decaying tail of thousands of diffuse reflections. Modern AI de-reverberators — including Clarity Vx DeReverb Pro and the machine-learning models inside RX's De-reverb module — are trained to separate direct sound energy from reflected energy using temporal patterns. When a noise floor is present, the reflected energy and the noise occupy overlapping spectral regions, and the model's confidence in distinguishing 'tail' from 'noise' drops sharply. The result is either incomplete reverb removal or audible artifacts: chirping, smearing, and a hollow, underwater quality.

The reverse order fails for a different reason. If you de-reverb a noisy file first, the reverb removal process boosts and shapes residual noise along with the signal, effectively 'baking' the noise into the processed audio. Your subsequent de-noise pass then has to work harder, requiring more aggressive reduction settings, which produces more artifacts. In short: each stage of processing degrades the material slightly, so you want the most forgiving, most accurate stage — noise reduction on stationary content — to happen while the signal is still as close to its original state as possible.

## The Practical Processing Chain: Step-by-Step

Here is the workflow that professional dialogue editors and podcast producers converge on, whether working in iZotope RX, Adobe Podcast Enhance-style web tools, Audacity with plugins, or an integrated AI toolbox like the one we build at audobox.com. Start by importing your audio at its native sample rate and bit depth — never resample before restoration, because each conversion adds quantization noise. Listen critically on headphones and identify your problems in priority order: hum, hiss, room tone, echo, plosives, mouth clicks, clipping.

Step one is always low-hanging fruit: remove tonal noise first if present. A 60 Hz hum with harmonics at 120, 180, and 240 Hz should be addressed with a notch filter or a dedicated De-hum module before anything else, because hum is perfectly periodic and trivially removable when isolated, but becomes entangled with musical content after other processing. Step two is broadband de-noise: capture a noise print from a section of pure room tone (ideally 1–2 seconds), set reduction between 6 and 12 dB initially, and increase gradually. Industry experience suggests keeping total noise reduction below roughly 15 dB per pass to avoid artifacts; beyond that, split the work across two gentler passes.

Step three is de-reverb. Apply it conservatively — most modern AI de-reverbers expose a single amount control, and values above 60–70% typically introduce noticeable smearing on speech. Step four handles remaining issues: de-click for mouth noises, de-plosive for pops, and finally corrective EQ and compression. A useful rule of thumb attributed to veteran post engineers: fix the environment before the performance, and the performance before the mix. Every stage downstream of restoration assumes the audio beneath it is clean.

## Comparison Table: De-Noise vs De-Reverb Characteristics

| Feature | De-Noise (first) | De-Reverb (second) |
| --- | --- | --- |
| Target signal | Stationary broadband noise (hiss, hum, fans) | Time-varying reflections and echo tails |
| Algorithm type | Spectral subtraction / noise profiling | AI source separation / transient analysis |
| Typical reduction range | 6–15 dB per pass, up to 20 dB with care | 30–70% wet/dry balance adjustment |
| Artifact risk | Musical noise, warbling at high settings | Smearing, hollow tone, loss of ambience |
| Sensitivity to input quality | High — needs clean noise profile | Very high — noise confuses the model |
| Best-in-class examples | RX Spectral De-noise, Lalal.ai, BandLab Voice Cleaner | Clarity Vx DeReverb Pro, RX De-reverb, Dialogue Isolate |
| Ideal position in chain | Early, right after de-hum | After de-noise, before EQ/compression |
| CPU cost | Moderate | Higher (neural inference) |

## Alternatives and When to Break the Rule
Not every situation fits the standard order, and pretending otherwise would be dishonest. Consider these documented exceptions. First, extreme room echo: if a recording was made in a bathroom or stairwell where the direct-to-reflected ratio is terrible, applying noise reduction first can exaggerate the smear. Some engineers run a light de-reverb pass at 20–30% just to tighten the signal, then de-noise, then finish with a second, stronger de-reverb pass. This two-pass sandwich approach trades processing time for artifact control and works well on difficult location dialogue.

Second, AI all-in-one enhancers change the calculus. Tools like Adobe's Enhance Speech, BandLab's Voice Cleaner (covered by MusicTech), and several entries in Unite.AI's August 2026 roundup of AI audio enhancers process noise and reverb jointly inside a single neural network. With these, there is no user-controlled order — the model was trained end-to-end on mixed degradation, and it decides internally. That can be convenient, but it also means less control; if the result sounds over-processed, your only lever is dialing back the enhancement strength rather than adjusting individual stages. Third, music sources behave differently than speech: reverberation on sung vocals is often musically intentional, and removing it can flatten a performance. For music, aggressive de-reverb is usually the wrong tool entirely, and noise reduction should be equally restrained.

Finally, consider whether you need these processors at all. Prevention remains cheaper than cure: a $100 dynamic microphone like a Shure SM58 or MV7 placed 10–15 cm from the speaker's mouth, plus a duvet or closet full of clothes as an improvised absorber, eliminates most of the problems that make people reach for de-reverb in the first place. Six Colors' walkthrough of cleaning up podcast room noise, hum, and echo makes exactly this point — the best restoration pass is the one you never need.

## Common Mistakes That Ruin Audio Restoration

The most frequent error, by a wide margin, is over-processing. Creators stack a de-noiser at maximum strength, then a de-reverberator at maximum strength, then an AI enhancer on top, and wonder why their voice sounds like it was transmitted through a tin can underwater. Each pass removes not just the unwanted content but also low-level signal detail — breath, consonant transients, room character — and the compounding damage is irreversible. A better discipline: apply the minimum effective amount at each stage, audition with A/B bypass comparison, and stop as soon as the problem stops being distracting rather than chasing absolute silence.

The second common mistake is skipping the noise profile step. Spectral de-noise tools that rely on a learned noise print produce far worse results when fed a generic preset instead of a profile captured from actual room tone in the same recording. Take the extra thirty seconds to find a gap where nobody speaks. Third, people often de-reverb content that does not need it — a small amount of natural room reflection makes dialogue sound human, and stripping it entirely creates the uncanny, dead 'AI voice' sound that listeners immediately notice. Aim for naturalism, not sterility. Fourth, avoid processing MP3-compressed source files more than necessary; lossy compression artifacts interact badly with both noise reduction and de-reverb models, so if you have access to WAV originals, always restore those instead. Fifth, do not forget gain staging: restoring a recording that was recorded 30 dB too quiet amplifies the noise floor along with everything else, so normalize gently (around −18 dBFS average) before restoration begins.

## Cost, Tools, and What You Should Expect to Pay

The price spread across this category is enormous, and expensive does not automatically mean better for your use case. At the free end, Audacity offers competent noise reduction via noise profiling, and Adobe offers free web-based speech enhancement with limits. BandLab's Voice Cleaner is free within its platform and handles both noise and reverb reasonably well for spoken word, as MusicTech's testing showed. Subscription AI services like Lalal.ai operate on a pay-per-minute or tiered basis, typically ranging from about $15 to $35 per month depending on volume.

At the professional end, iZotope RX 11 Standard lists around $399 and RX Advanced around $1,199, with upgrade pricing significantly lower for existing users — the RX line has been the industry standard for audio repair since version 1 shipped in 2007, and the RX 12-era feature set reviewed by ProVideo Coalition continues that lineage. Waves Clarity Vx DeReverb Pro sells for around $299 list but frequently appears in Waves' perpetual sales at $49–$99, and its TEC Award recognition reflects genuine quality on dialogue-heavy material. For creators who want capable results without assembling a plugin arsenal, integrated AI toolboxes — audobox.com among them — bundle enhancement, cleanup, and generation into one subscription, typically in the $10–$25 per month range. The honest assessment: free tools now cover perhaps 80% of hobbyist needs, mid-tier subscriptions cover 95% of creator needs, and RX Advanced earns its price mainly for film/TV post-production professionals working against deadlines on compromised location audio.

## When to Act: Fixing It Now vs Re-Recording

Restoration has limits, and knowing when to stop processing and simply re-record is a mark of competence, not failure. As a threshold guide: if your signal-to-noise ratio is worse than roughly 20 dB (meaning the noise is clearly audible during pauses and even during speech), or if reverberation time exceeds about 0.8 seconds in the recording space, no software will fully rescue the take. AI tools have improved dramatically — Dialogue Isolate-class systems can pull usable dialogue out of remarkably bad recordings, and the 2025–2026 generation of enhancers handles cases that were hopeless five years ago — but 'usable' and 'broadcast-quality' remain different things.

Act immediately when the problem is mechanical and cheap to fix: re-record a voiceover in a treated corner rather than spending two hours fighting a stairwell echo. Act with software when re-recording is impossible — interviews with people who are unavailable, archival footage, live event recordings, or client-supplied audio. And act early in your project timeline: restoration should be the first stage of post-production, before editing decisions lock in processed audio. One practical tip from working editors: keep unprocessed backups of every file. Restoration algorithms improve yearly, and a file you salvage at 85% quality today might hit 95% with next year's tools — but only if you preserved the original.

## Final Recommendations by Use Case

For podcasters, the prescription is simple: record in the quietest space available, get close to a dynamic mic, then run de-hum (if needed), de-noise at moderate settings, and only touch de-reverb if the room genuinely echoes. Free tools like BandLab Voice Cleaner or Adobe's web enhancer will handle most episodes; upgrade to RX Standard or a paid AI toolbox when clients demand consistency across dozens of episodes. For video creators and filmmakers, invest in RX or Clarity Vx DeReverb Pro, because location dialogue is where professional-grade restoration pays for itself fastest — a single rescued interview can justify the purchase. For musicians, be conservative: noise reduction on a full mix is risky, and de-reverb on intentionally reverberant vocals usually damages the aesthetic. Restore individual stems or raw tracks before mixing instead.

Across all cases, remember the core principle this article established: de-noise before de-reverb, de-hum before both, and restraint everywhere. The order matters because each algorithm's assumptions break down when fed the output of the other, and the difference between a clean, natural-sounding result and a smeared, robotic mess often comes down to nothing more exotic than sequence and moderation.

## Quick answers

### Can I run de-reverb and de-noise at the same time?

Some AI tools like Dialogue Isolate and Adobe Enhance process both simultaneously inside one neural model, which works well for speech. However, with traditional modular plugins, running them in parallel on the same track causes unpredictable interaction — serial processing with de-noise first gives you more control and cleaner results.

### How much noise reduction is too much?

Most engineers recommend staying below 12–15 dB of reduction per pass. Beyond that, spectral subtraction starts producing 'musical noise' — watery, chirping artifacts — and AI tools begin smearing consonants. If you need more, use two gentle passes rather than one aggressive one.

### Does de-reverb work on music or only speech?

De-reverb tools are trained primarily on speech and work poorly on full music mixes, where reverb is often intentional. On solo vocals or dry instrument stems they can help, but expect artifacts. For music, it's better to fix the recording space or adjust reverb sends in the mix.

### Is iZotope RX worth it compared to free tools?

RX Standard (~$399) is worth it for professionals handling compromised location audio regularly, thanks to modules like Spectral De-noise and De-reverb with fine manual control. For podcasts and casual projects, free options like Audacity's noise reduction or BandLab's Voice Cleaner cover most needs adequately.

### What's better: preventing noise or removing it later?

Prevention wins almost every time. A dynamic microphone used close-up in a soft-furnished room eliminates most noise and echo problems at the source, preserving audio quality that no restoration can recover. Software cleanup should be a safety net, not a substitute for decent recording technique.

Canonical: https://audobox.com/knowledge/should_you_de-reverb_or_de-noise_first.php
Markdown: https://audobox.com/knowledge/should_you_de-reverb_or_de-noise_first.php/index.md