# Review AI Speech Cleanup Against the Original

> Compare an AI-cleaned voice recording with the original using timed listening notes, intelligibility checks and a clear keep, retry or rerecord decision.

Compare an AI-cleaned voice recording with the original using timed listening notes, intelligibility checks and a clear keep, retry or rerecord decision.

Review AI speech cleanup by comparing the processed recording with the preserved original at the same intervals. Record what a listener heard in both versions, then decide whether to use the processed interval, keep the original, retry, rerecord, or leave it pending. Quieter background is only one observation; the words and pronunciation still need checking.

This workflow produces a listening review sheet and a list of unresolved windows. It does not process audio or assemble a master. Those actions follow the user's decisions through the chosen editing tool.

## Preserve the recording you will compare against

Keep the original file unchanged and name each processed variant distinctly. Record the original ID, variant ID, duration, and acceptance criteria. For example, your criteria might require preserving every spoken word and a location name that is essential to the interview. Treat those criteria as the user's priorities, not a universal audio-quality score.

[Adobe Podcast](/tools/adobe-podcast) provides Enhance Speech, which filters noise and artifacts; Adobe notes that results depend on the original audibility and background noise. [Its FAQ](https://helpx.adobe.com/podcast/adobe-podcast-faq.html) describes those limits.

Adobe's Enhance Speech workflow accepts uploads and provides a processed download. Mic Check addresses recording setup before capture. [Adobe's Enhance Speech introduction](https://podcast.adobe.com/guides/what-is-enhance-speech) describes these distinct uses.

Do not overwrite the source when trying a processed variant. If a recording lacks a recoverable word, cleaner output is not proof of what the speaker intended. Keep the comparison anchored to observable differences rather than asking the model to invent a restoration.

## Choose windows the listener can identify

Use start and end times from the same recording timeline. Mark intervals around concerns such as a changed consonant, an obscured word, a speaker transition, or a noise-heavy phrase. Include enough surrounding speech for the listener to judge the word in context.

Record the interval basis explicitly. If the processed file includes a trim or moved section, establish how its timeline maps to the original before comparing it. Matching numbers are not enough if they refer to different content. With missing duration or alignment information, return the gap instead of pretending the intervals are bounded.

Read only the supplied notes and recording identifiers. A filename does not grant access to a media player, and a file-reading tool does not audition a WAV. If no supported listening surface or human observations are available, provide a blank review template and request the paired observations needed to fill it.

## Work from fictional paired listening notes

This neighborhood interview is fictional. The files are `bench-interview-original.wav` and `bench-interview-clean-v1.wav`. Listener J. Chen compared both. The observations below are supplied human notes; an assistant did not hear the files through text access.

| Segment | Window | Original observation | Processed observation | Listener decision |
| --- | --- | --- | --- | --- |
| A1 | 12–17 s | The phrase contains “the west bench.” | The phrase sounds like “the wet bench.” | Keep original: the location word changed. |
| A2 | 40–45 s | Traffic is audible behind the speech. | Traffic is quieter; all spoken words are retained. | Use processed: listener confirmed the words remain intact. |

The review sheet should also record the listener, reason, and whether the comparison was actually performed. Those facts are declared review evidence, not something a checker can independently observe. Do not copy “listened: true” into unreviewed rows just to make a packet pass.

## Decide what improved and what was lost

For A1, the important comparison is west versus wet. Even if other noise decreased, the supplied note identifies a change to an intelligible location word. The listener's keep-original decision preserves that word for this window. Retain both observations so a later editor can understand the choice.

For A2, the listener confirmed reduced traffic and retained words, so the processed interval is an accepted candidate. That decision applies to A2, not automatically to the entire recording. Review the remaining windows before treating the variant as a complete master.

[Speech enhancement](/glossary/speech-enhancement) names the processing task. This review adds a separate acceptance decision. If processing helps the background but obscures speech, document both results and let the listener choose between another attempt, the original, a new recording, or a pending decision.

## Give incomplete rows a real status

The [Speech Cleanup Review Sheet](/skills/marketing/speech-cleanup-review-sheet) organizes segment IDs, millisecond bounds, paired observations, decision, reason, listener, and `listened` state. Keep an interval pending if either observation is absent. Without observations, a transcript or waveform appearance should not be presented as a completed listening comparison.

Use the exact machine decision labels when preparing the companion `audio-review-sheet.json`: `keep-original`, `use-processed`, `retry`, `re-record`, or `pending`. Include the recording's real `duration_ms`; do not infer it from the last listed review interval. Millisecond numbers should be integers on the agreed timeline.

A retry is a proposed next action, not evidence that a better result exists. A rerecord decision should say which phrase needs another capture. A pending row should identify who will listen and what uncertainty they need to resolve. This keeps the sheet useful even when there is no accepted replacement.

## Use code to find bookkeeping gaps

The companion checker reads the fixed review-sheet basename without arguments in a Python 3 POSIX environment. It flags intervals outside the declared duration, missing paired observations, pending decisions, and accepted decisions lacking declared listener evidence or a reason. It emits a JSON result with a structural status.

It cannot hear pronunciation, assess a noise reduction, or verify that the named listener performed the comparison. A successful result means those declared fields and bounds meet its checks. Retain the human comparison as the basis for choosing audio.

## Review the eventual master at its transitions

After the user applies accepted choices through an editing tool, have a listener review the assembled master. A1 and A2 may be reasonable independently while a join between different variants needs adjustment. Compare the master with the original where an edit affects wording, timing, or meaningful environmental sound.

Finish with the preserved original, named processed variants, completed review rows, and explicit pending work. The listener approves words, pronunciation, and meaningful sound before choosing the final master. Keep upload, processing, overwriting, and publication outside these review installables; their output is the decision record that guides the user's next action.

## Sources and further reading

- [Adobe Podcast FAQ](https://helpx.adobe.com/podcast/adobe-podcast-faq.html) — Adobe
- [What is Enhance Speech?](https://podcast.adobe.com/guides/what-is-enhance-speech) — Adobe

---

_Source: https://agentscamp.com/guides/voice/review-ai-speech-cleanup — Guide on AgentsCamp._