> ## Content Index
> Fetch the complete content index at: https://izotoperx.net/llms.txt
> Use this file to discover other available public pages before exploring further.

# iZotope RX 12 Spectral Recovery: Fix Thin Audio
- URL: https://izotoperx.net/izotope-rx-spectral-recovery/
- Published: 2026-09-04T10:25:36.000Z
- Updated: 2026-09-08T09:31:02.000Z
- Description: Use Spectral Recovery only where the spectrogram shows a real bandwidth cutoff, then add the smallest believable low and high extension without inventing brittle speech.
- Author: Brandon Hayes
- Tags: Repair Modules, Dialogue Cleanup, RX 12

**Use RX 12 Spectral Recovery only when speech has a real frequency ceiling or missing bands.** Confirm the cutoff in the spectrogram, repair any dominant clipping, noise or isolated codec damage, let Learn suggest the low and high boundaries, then add the smallest believable extension. Stop when clarity becomes lisping, brittle, phasey or boomy.

Spectral Recovery synthesizes plausible frequency content. It does not retrieve the speaker's original missing harmonics from a phone call. The professional goal is a more intelligible, less fatiguing match to the surrounding production—not a fake studio microphone.

## Spectral Recovery at a glance

| Problem                                         | First tool                       | Why                                        |
| ----------------------------------------------- | -------------------------------- | ------------------------------------------ |
| Hard high-frequency cutoff in phone/VoIP speech | Spectral Recovery                | Synthesizes energy above the real boundary |
| Missing/filtered natural low end                | Spectral Recovery                | RX 12 can generate low-frequency extension |
| Small codec holes between boundaries            | Spectral Patching                | Fills gaps from surrounding spectrum       |
| Steady hiss or fan                              | Voice/Spectral De-noise          | Noise reduction, not bandwidth extension   |
| Changing noise and room reverb                  | Dialogue Isolate                 | Separates voice, noise and reverb controls |
| Packet dropout or one codec chirp               | Spectral Repair/dropout workflow | Local missing event, not a global cutoff   |

## What RX 12 Spectral Recovery does

The [current RX 12 manual](https://docs.izotope.com/rx12/en/spectral-recovery.html?ref=izotoperx.net) says Spectral Recovery automatically adds missing frequencies to bandwidth-limited speech and patches holes caused by compression artifacts. It is designed for sources such as Skype or Zoom recordings with a hard upper cutoff and potentially reduced natural low end.

The module is included only in RX 12 Advanced, according to the [official edition comparison](https://www.izotope.com/products/rx-advanced?tab=compare&ref=izotoperx.net). It runs as a module in the standalone Audio Editor, not as a Spectral Recovery AU, VST3 or AudioSuite plug-in. Standard and Elements do not include it. Open the protected audio file in Advanced and find Spectral Recovery in the module list.

## Diagnose bandwidth loss before processing

Open the spectrogram and find sustained speech harmonics. A true upper limit appears as repeated voice content that stops abruptly at roughly the same frequency while the area above remains empty or carries only codec noise. A reduced low end may appear as filtered lower harmonics, but a quiet lower spectrum alone is not proof of damage: speaker pitch and microphone response matter. Inspect several voiced phrases and listen before deciding that bass needs synthesis.

Do not diagnose by “thin” alone. A distant microphone, comb filtering, aggressive noise suppression, excessive high-pass filtering, room reverb and phase cancellation can all sound thin. Spectral Recovery treats missing bands; it does not correct every cause of weak tone.

If the display is unfamiliar, use the [RX spectrogram guide](https://izotoperx.net/how-to-read-izotope-rx-spectrogram/). Mark the observed upper and lower edges before pressing Learn so the suggestion can be checked rather than accepted blindly.

## Synthesized detail is not recovered original detail

The module predicts content that fits the remaining speech. It cannot know the exact air, sibilance, microphone response or room information that the codec discarded. Two plausible restorations can sound different while neither is the lost original.

Use honest language in a production note: “bandwidth-extended phone recording,” not “restored full-resolution master.” Keep the original clip available for legal, documentary and editorial reference. If the call sound is part of the story, partial improvement may be more truthful than making it unnaturally polished.

## Repair the problems that would mislead recovery

If clipping is present, repair it before judging low/high extension; distortion can obscure the speech cues you are trying to assess. The [De-clip workflow](https://izotoperx.net/fix-clipped-audio-rx-de-clip/) covers that repair. Remove severe hum and isolated digital spikes before synthesis so the added bands do not make those defects more conspicuous.

When noise or reverb dominates, reduce it conservatively first. iZotope's [phone and Zoom cleanup guide](https://www.izotope.com/community/blog/how-to-clean-up-audio-from-phone-or-zoom-recording?ref=izotoperx.net) is a September 2022 walkthrough: its Zoom example uses EQ and Dialogue De-reverb before Spectral Recovery. Its tools overview separately recommends Spectral De-noise or Dialogue Isolate for noise. This is a useful example of sequencing, not a mandatory chain for every call. Avoid heavy reduction that erases the speech cues the recovery model needs.

## Build a diagnostic speech selection

Choose a short but representative passage with vowels, S/F/SH sounds, hard consonants, quiet words and the loudest phrase. Let the variety of speech, rather than a fixed number of seconds, determine its length. Include damaged phrases and room tone in the evaluation, while keeping a cleaner representative speech selection for Learn. A sustained vowel may make the extension sound smooth while the next sibilant turns artificial.

Keep a duplicate file or RX Document, then use the same pre-roll for every comparison. The [Preview and Compare workflow](https://izotoperx.net/izotope-rx-selections-preview-compare-history/) helps keep boundary and gain changes auditable. Match apparent loudness and replay the same phrase; louder or brighter is not automatically more intelligible. The [current module-controls guide](https://docs.izotope.com/rx12/en/common-module-controls.html?ref=izotoperx.net) documents Preview, Bypass and Compare. Compare stores named candidates for audition; it does not replace saving the approved audio.

## Use Learn as a suggested map

Learn analyzes the selection and suggests Low Cutoff and High Cutoff values. Compare them with the visible spectrum. A mixed passage with music, notification tones or a second microphone can confuse the boundary, so learn from isolated representative speech where possible.

Do not treat Learn as an approval. Move each cutoff around the actual edge and preview complete phrases. If the suggested low boundary falls inside speaker rumble or the high boundary sits below useful recorded consonants, correct it manually.

## Set Low Cutoff and Low Gain

Low Cutoff defines the frequency below which Spectral Recovery synthesizes low-frequency signal. With Spectral Patching enabled, holes between this boundary and High Cutoff can also be filled. Raising Low Cutoff expands the range assigned to low-frequency synthesis; it is not a high-pass filter that removes bass. The manual suggests trying a cutoff just above the fundamental frequency of the voice. Treat that as a starting experiment for VoIP speech, not a fixed value for every speaker. If the fundamental is missing, do not pretend the empty region reveals its exact original level.

Low Gain controls the synthesized low-frequency level. In the current module illustration it is the Gain slider to the left of the frequency graph; the right-hand Gain slider controls the high extension. Add only enough body to make the voice sit with adjacent dialogue. Stop when plosives swell, room rumble appears, the speaker becomes chesty or pitch feels unstable. Small phone speakers may not reveal the error, so check full-range monitors and headphones.

## Set High Cutoff and High Gain

High Cutoff defines the boundary above which high-frequency content is synthesized. With Patching enabled, gaps between Low Cutoff and High Cutoff may be filled. Lowering High Cutoff extends synthesis farther into the upper speech band, so do not move it down across useful recorded consonants merely to get more brightness. Put it at the real recorded ceiling, not at a round number copied from a tutorial. Different codecs and meeting modes create different limits. The 4,000 Hz value in the vendor illustration is an example, not a requirement for every phone or Zoom file. Zoom’s [current codec documentation](https://library.zoom.com/admin-corner/network-management/quality-of-service-and-network-best-practices-explainer/how-zoom-audio-and-video-codecs-affect-bandwidth-usage?ref=izotoperx.net) describes adaptive Opus and SILK audio; the platform name alone does not establish a fixed recorded cutoff.

High Gain sets the synthesized upper energy. Raise it until consonants become easier to follow, then compare one step lower. Stop at lisping, spitty S sounds, metallic air, phasing or a detached “spray” above the voice. iZotope's [2022 phone and Zoom walkthrough](https://www.izotope.com/community/blog/how-to-clean-up-audio-from-phone-or-zoom-recording?ref=izotoperx.net) warns that too much Spectral Recovery can create artifacts and shows a lisping example.

## Use Spectral Patching for codec holes

When enabled, Spectral Patching fills holes between the low and high cutoffs by sampling the area around the missing audio. The manual notes that the effect can be subtle by ear even when the spectrogram looks more continuous. Do not approve it because the picture is prettier.

Toggle it at matched loudness on words crossing visible holes. If consonants smear or codec movement becomes more obvious, leave it off and repair the isolated event with the [Spectral Repair workflow](https://izotoperx.net/izotope-rx-spectral-repair/). Patching is not packet-loss reconstruction for long missing syllables.

## Avoid controls from older Spectral Recovery tutorials

The [RX 9 manual](https://s3.amazonaws.com/izotopedownloads/docs/rx9/en/spectral-recovery/index.html?ref=izotoperx.net) shows Amount, Vowel/Sibilant Balance and Smoothing, controls also associated with the earlier RX 8 design. The RX 12 manual documents Learn, Spectral Patching, Low Gain, Low Cutoff, High Cutoff, High Gain and Render. The low-frequency extension was already documented in RX 10; its presence is not evidence of a newly added RX 12 feature. If the screen in a tutorial does not match yours, check its RX version before copying settings.

This matters for search results: older advice may be conceptually useful but cannot be a current click-by-click guide. Our [RX 12 changes guide](https://izotoperx.net/izotope-rx-12-new-features/) explains the active release context.

## Keep noise, reverb and recovery as separate decisions

Steady hiss belongs in Voice or Spectral De-noise; changing noise and reverb may belong in Dialogue Isolate. Follow the [denoise comparison](https://izotoperx.net/voice-denoise-vs-spectral-denoise/) and [Dialogue Isolate tutorial](https://izotoperx.net/izotope-rx-dialogue-isolate/). Judge the generated band directly: a fuller spectrum is no guarantee that the surviving noise or codec texture will sound natural.

After recovery, listen again to room tone. Synthetic high energy can make a previously tolerable codec bed feel fizzy. If a second gentle noise pass is required, compare it with a version that used less High Gain; two aggressive processes are usually worse than one restrained compromise.

## EQ and resampling do not recreate a missing band

EQ can shape recorded frequencies, but boosting above a hard cutoff only boosts noise or nothing. Resampling changes the file's sample grid; it does not invent the discarded harmonics. Use EQ for broad tonal balance and Spectral Recovery only for the missing-band problem it was built to address.

A gentle EQ after recovery may seat synthetic content with production dialogue, but first solve cutoff and gain inside the module. Otherwise the EQ can hide a bad setting that returns on another speaker or playback system.

## Remote interview workflow

Use isolated participant tracks whenever possible. Repair each speaker separately, because one local microphone may be full-band while the call mix is limited. If only the compressed mix survives, select regions carefully so hold music and notifications do not enter Learn.

The current iZotope [audio-cleanup guide](https://www.izotope.com/community/blog/how-to-clean-up-audio-and-remove-background-noise?ref=izotoperx.net) identifies Spectral Recovery as a route for hollow dialogue. The complete [podcast cleanup workflow](https://izotoperx.net/izotope-rx-podcast-cleanup/) handles clipping, noise, clicks, plosives, levels and export around it.

## Look for a better source before synthesizing anything

Ask whether the conferencing platform stored isolated local tracks, whether the guest recorded a phone voice memo, or whether a camera and screen recorder captured alternate audio. A compressed call mix may have a hard ceiling while the local backup retains the original microphone bandwidth. Syncing a better source restores access to recorded detail; Spectral Recovery still estimates missing content. Verify timing and drift across the whole replacement.

Check messaging attachments and cloud previews carefully. A service may create a compressed listening proxy while the downloadable original is better. Preserve both and document which file entered the repair. Our [RX import, export and safe-saving guide](https://izotoperx.net/izotope-rx-import-export-save/) covers version separation and format checks.

For future remote sessions, arrange a local recording for each participant, use headphones, and test the platform’s processing settings before the call. Keep a sync slate and the call recording as a reference. An [April 2022 AudioPost discussion about swimmy Zoom sound](https://www.reddit.com/r/AudioPost/comments/u2w84y/?ref=izotoperx.net) highlights the practical distinction: changing capture settings may help a future call, but does not repair the recording already delivered. Its RX 8/9 suggestions are historical experience, not an RX 12 quality benchmark. No restoration module beats receiving the microphone signal before the meeting codec removed it.

## Do not confuse bandwidth loss with packet loss

A bandwidth cutoff is persistent: speech repeatedly ends at the same upper or lower boundary. Packet loss can produce short dropouts, freezes or repeated fragments. Warbling by itself does not prove packet loss; codec compression and noise suppression can sound similar. Spectral Recovery can extend the surviving band but cannot reconstruct a long missing word from nothing.

For a short interruption, inspect the event and try the [RX audio-dropout workflow](https://izotoperx.net/izotope-rx-repair-audio-dropout/) or a local Spectral Repair selection. For a missing phrase, patch from another take, an isolated participant track or an editorial pickup. Do not cover missing language with synthetic brightness.

iZotope’s [February 2021 compressed-interview guide](https://www.izotope.com/community/blog/tips-to-repair-a-compressed-or-noisy-interview?ref=izotoperx.net) treats noise, clipping and codec damage as separate defects. Follow that diagnostic separation even when several problems occur in one remote clip.

## Match the production, not an imaginary full-band voice

Place the repaired clip between the lines that precede and follow it. Compare apparent distance, brightness, room, intelligibility and background. A little remaining telephone character can sound natural in a documentary; an overextended voice can jump forward and make the edit obvious.

If the target is an audiobook pickup or a host insert, match the approved narrator master rather than maximizing bandwidth. The [audiobook cleanup and delivery workflow](https://izotoperx.net/izotope-rx-audiobook-cleanup/) adds chapter continuity, RMS, peak and noise-floor gates. If you are deciding whether this Advanced-only module justifies an upgrade, use the [RX edition guide for podcasters](https://izotoperx.net/best-izotope-rx-edition-podcasting/).

The official [RX 10 Spectral Recovery demonstration](https://www.youtube.com/watch?v=mGpojGVcg8k&ref=izotoperx.net) is another historical tutorial reference. Its settings belong to its example source; use the current RX 12 manual for today’s controls. Your own codec boundary and speaker determine the usable values.

Use a small corrective EQ only after the extension survives direct comparison. Bypass both EQ and recovery independently so you know which stage improves matching. Save one conservative and one stronger candidate, then choose by intelligibility, artifacts and scene continuity. Do not infer a client preference from how impressive the solo example sounds.

## Render, label and hand off the extended clip safely

Work on a duplicate and retain useful handles. Render applies the module to the current selection in the open editor file; then save or export the approved result under a new filename. It does not automatically create a separate deliverable. Record RX version, Low/High Cutoff, Low/High Gain, Patching state and the processed range. Label the asset `bandwidth-extended` so another editor does not mistake it for a full-band original.

After export, reopen the actual file and inspect its spectrum. A new codec can remove part of the synthesized band or add pre-echo, warble and inter-sample peaks. Listen at the destination data rate and on the devices viewers will use. A good lossless editor preview is not evidence that a streaming encode survived. Check the exported duration, channel layout and head/tail alignment against the source; listen for artifacts throughout every processed region, not just the test phrase.

The [May 2026 iZotope spoken-word workflow](https://www.izotope.com/en/learn/fast-audio-cleanup?ref=izotoperx.net) places restoration before level and delivery control. Keep that dependency: any later denoise, EQ, limiter or encoder can change the artifact balance. Approve the final deliverable, not the intermediate module output.

For formal review, retain the unprocessed original, a repair-only version and the final mix. A reviewer can then distinguish source damage from the choices made in post. For archival, journalistic or research material, keep the untouched recording alongside any listening copy and document the processing. Synthetic detail should not be treated as newly recovered evidence of what was said.

## Phone and video-call matching

Compare against the scene's other dialogue, not a commercial voiceover. A phone character may need to remain narrower to make narrative sense. Add enough extension for intelligibility and consistency, then preserve the recognizable source perspective.

If the call alternates speakers with different codecs, split them into regions and learn/test independently. A single High Cutoff may begin synthesis inside one speaker’s useful recorded band while starting too high to address another’s lower ceiling. Check each speaker and any change in connection quality separately.

## Artifact diagnosis

| Artifact                     | Likely cause                                          | First move                                    |
| ---------------------------- | ----------------------------------------------------- | --------------------------------------------- |
| Lisping or spitty consonants | High Gain too high or bad High Cutoff                 | Lower gain and recheck actual ceiling         |
| Detached metallic “air”      | High extension too prominent                          | Reduce High Gain; compare in context          |
| Boom/chestiness              | Low Gain too high                                     | Reduce Low Gain; recheck the learned boundary |
| Unstable bass pitch          | Low synthesis has weak cues                           | Use less low extension                        |
| Fizzy background             | Noise or codec texture becomes more audible           | Control noise or use less High Gain           |
| Smeared words                | Possible patching artifact or excessive prior denoise | Disable Patching and compare lighter repair   |

## Ten-step Spectral Recovery workflow

1. **Protect the source.** Duplicate the original phone, meeting or archive recording.
2. **Confirm a real bandwidth problem.** Separate a hard cutoff from noise, reverb, clipping and poor mic position.
3. **Repair blocking defects first.** Correct clipping, severe hum, changing noise and isolated codec clicks.
4. **Choose representative speech.** Include vowels, sibilants, quiet/loud words and the worst damage.
5. **Use Learn as a starting point.** Compare suggested cutoffs with the visible boundaries.
6. **Set Low Cutoff and Low Gain.** Add body without boom, rumble or pitch instability.
7. **Set High Cutoff and High Gain.** Add clarity without lisping, fizz or brittle consonants.
8. **Test Spectral Patching.** Use it only for visible compression holes that improve by ear.
9. **Compare at matched loudness.** Check phrases, room tone, speakers and headphones.
10. **Render and verify delivery.** Apply to the protected selection, save a new version and listen through every processed region in the final encoded file.

## Quality-control checklist

- The spectrogram shows a real repeatable bandwidth boundary.
- Learn values were checked against the source, not accepted automatically.
- Quiet words, vowels and every major sibilant were auditioned.
- Low extension does not add boom, rumble or unstable pitch.
- High extension does not add lisping, metal, fizz or detached air.
- Patching improves audio, not just the spectrogram.
- Comparison is matched in level and uses the same pre-roll.
- Original and extended versions remain clearly labeled.
- Final export preserves sync, channel layout and intended perspective.

If you are evaluating Advanced, the [iZotope demo policy](https://support.izotope.com/hc/en-us/articles/6658022746257-How-to-Demo-iZotope-Software?ref=izotoperx.net) disables saving and exporting in standalone RX Trial and Demo modes. Test a short passage in-app before buying; a rendered preview in that mode is not a deliverable you can export.

The [Repair Audio hub](https://izotoperx.net/repair-audio-rx/) connects Spectral Recovery with De-clip, Spectral Repair, dropout repair and other source-specific restoration.

## What not to do

- Do not run Spectral Recovery on full-band studio speech as an enhancer.
- Do not describe synthesized content as the original recovered signal.
- Do not copy RX 8/9 controls into an RX 12 workflow.
- Do not let Learn replace visual and listening diagnosis.
- Do not maximize High Gain for the brightest solo result.
- Do not use Low Gain to compensate for unrelated microphone tone.
- Do not judge Patching from the spectrogram alone.
- Do not overwrite the source or skip encoded-file playback.

## Frequently asked questions

### What does RX Spectral Recovery do?

It synthesizes missing low and high frequencies in bandwidth-limited speech and can patch smaller spectral holes caused by compression artifacts. It is designed for phone, VoIP, meeting and similar recordings.

### Which RX 12 edition includes Spectral Recovery?

Spectral Recovery is included in RX 12 Advanced. The current official comparison does not list it in Standard or Elements.

### Does Spectral Recovery restore the original missing frequencies?

No. It generates plausible content from the surviving speech. The result can sound fuller or clearer, but the discarded original information is not recovered exactly.

### What does Learn do in RX 12 Spectral Recovery?

Learn analyzes the selection and suggests Low Cutoff and High Cutoff values. Check both against the spectrogram and adjust them for the actual speaker and codec.

### What does Spectral Patching do?

It fills holes between the low and high cutoff frequencies by sampling surrounding spectrum. The change may look clearer than it sounds, so approve it by ear.

### Why does Spectral Recovery make speech lisp?

High Gain may be too strong, the High Cutoff may be misplaced, or the source lacks stable cues for sibilants. Lower the synthesized high level and compare full phrases.

### Should denoise come before Spectral Recovery?

Usually repair dominant noise or reverb conservatively first so recovery does not extend the defect. Avoid overprocessing that removes speech cues needed for synthesis.

### Can EQ or upsampling do the same job?

No. EQ reshapes recorded energy and resampling changes the sample grid. Neither recreates a missing band; Spectral Recovery synthesizes new content across diagnosed boundaries.