PixelOkay
● Local audio
Tools
PIXEL OKAY · AUDIO STUDIO

Podcast Cleaner Studio

Find fillers, repeated phrases, and awkward pauses. Review every suggested cut, fine-tune your transcript, and export audio without sending your recording to a server.

Free Whisper transcriptionEditable transcriptNon-destructive cutsLocal WAV export

THE WORKSPACE

Your editing desk

Choose a recording to get started.
01

Load your recording

Audio stays on this device.

♫Drop an audio file hereMP3, WAV, M4A, OGG and other browser-decodable audio
Microphone is off · permission requested only when recording starts
Drag on waveform to select; click to seek.

Preview plays the original until you select Preview edited.

DSP

No-AI Voice Cleanup

Real local filters; same chain in preview and WAV.

Audio-reactive DSP, not AI restoration. Expander may reduce quiet speech; compressor and peak ceiling may change dynamics. Always A/B listen before export. No original recording is changed.

02

Find edits

No-AI quiet-gap review first; optional AI transcript separately.

✓ No AI · Local audio editorImport → Clean → Edit → Listen → Export. No model or network required.
✧ Enable local AI transcription (optional)

Whisper models require an explicit download. Transcription is not necessary for waveform cutting, cleanup, or audio export.

AI not loaded

AI model downloads start only when you confirm Download & Load in AI Model Setup. Analyze opens setup if the selected model is not ready. Local inference can be slow on CPU. Imported word timestamps can be used without AI. Quiet-gap scanning uses signal amplitude, not speech recognition, and may flag softly spoken words.

03

Manual edit

Cut a specific time range.

Short crossfades help avoid clicks. Results are not equal to professional multitrack editing.

Review suggestions

Nothing is removed until you enable a cut.

0 suggestions
—Original duration
0:00Proposed removal
—Estimated remaining
✧

Suggested edits will appear here

Add a recording and run Analyze, or import timestamped transcript JSON.

Editable transcript

Select a word to correct its text and timing, or approve its audio removal.

0 words

Your transcript appears after recognition. Word timings are estimates and should be reviewed before exporting cuts.

Text/timing corrections update transcript and suggestions; they do not rewrite spoken words. To change the sound, enable a suggested cut, click a word and choose its cut control, or add a manual time range.

Listen & export

Review every cut before making your final WAV.

Original preview

MADE FOR THE DETAILS

How it works

01

Transcribe or import

Use the free Whisper model locally after the initial download, or import a timestamped word transcript.

02

Approve edits

Review filler words, adjacent repeated phrases, and long pauses. Suggestions are not applied automatically.

03

Preview and export

The tool stitches the surviving audio together with short crossfades and exports a real WAV file.

Are my recordings uploaded?

No audio is uploaded by this application. The free inference library and model files load from external sources on first use; the speech data is processed locally.

Is every detection accurate?

No. Word timestamps from speech recognition are estimates. Words like “like” may be natural speech. Listen and review each edit before exporting.

What are the limits?

Files up to 400 MiB, duration up to 15 minutes, stereo audio, and a 320 MiB decoded-audio budget. WAV export is serialized in bounded chunks to reduce peak memory; browser codec support and memory limits still vary. English tiny is the default; multilingual tiny is less accurate than larger models.

LOCAL AI SETUP

Set up AI on this device

First, a model needs to be loaded on your device

This tool uses local Whisper speech recognition. Choose English or multilingual, then explicitly download/load the free model. Your recording stays in your browser; model/runtime assets come from external CDNs.

  1. 1. Check device
  2. 2. Download / cache
  3. 3. Initialize selected AI
  4. 4. Transcribe audio
Whisper tiny (English or multilingual), Transformers.js 3.8.1 / WASM / q8. Download and working-memory requirements vary by model and browser; they are not live RAM readings.

Device & model resources

Estimates, not live memory
Device RAM / model memoryChecking…

Model memory usage is not exposed.

Browser website storageChecking…

Site storage is not RAM.

ONE ACTION · ALL RECOGNIZED AI MODEL DOWNLOADS

Free AI model storage across PixelOkay

Scan this website origin for recognized WebLLM, Transformers.js and other identifiable model caches from all PixelOkay tools on this exact origin. Other website data and databases are never automatically cleared.

Scan for safely recognized model caches before deletion.

Unrecognized Cache Storage, IndexedDB and OPFS are not removed.

Having trouble downloading or loading AI?

Try Refresh Storage & Cache, a stable network and Retry Setup. Cache cleanup is optional and cannot fix every GPU, CORS, network or worker failure. It can require other PixelOkay tools to re-download models.

Check device capabilities, then choose whether to load AI.
Currently loadedNone

Remove downloaded AI model caches?

Using AI Podcast Cleaner

Import a spoken-word recording, run a silence scan or optional transcription and review each proposed cut. Suggestions start unchecked; select only intended removals, edit transcript/timing as needed and compare before exporting. Keep the audio and project backup together.

Silence/filler/repeat suggestions can mark quiet speech or legitimate words and require review. Whisper downloads models/runtime only after you approve model setup. Word/caption timing is approximate, especially around crossfades; spelling edits do not synthesize new speech. File/duration/decoded-memory budgets apply. Project matching uses sampled fingerprints, not full-file cryptographic identity, and cannot replace keeping the original audio.

References: Transformers.js runtime

Review every podcast cut before exporting audio

Import a supported recording and use the RMS silence scan or optional Whisper transcription to find candidate edits. Suggestions start unchecked: listen to each proposed cut and approve it manually so quiet speech, breaths or intentional pauses are not removed by assumption.

Example workflow

  1. Keep the original recording and run a silence scan or explicitly load optional transcription.
  2. Review suggested cut points against the waveform and transcript, especially around quiet speech, crossfades and word boundaries.
  3. Approve only wanted edits, preview the result, and export WAV and any supported transcript or caption files.

Authored example: In a short interview, inspect a long pause before a guest response. Play the audio on both sides of the cut and leave it unchecked if the pause carries meaning or the guest starts speaking quietly.

Limits and model handling: Optional Transformers.js Whisper assets download after model setup is approved; the RMS silence scan is model-free. Transcript word and caption timing is approximate, and spelling edits do not synthesize replacement speech. Project matching uses sampled fingerprints, not full-file cryptographic identity; keep the original recording. File, duration and decoded-memory budgets apply.

AI Podcast Cleaner questions

Are silence and transcript cut suggestions applied automatically?

No. Suggestions start unchecked and need your review and approval before they become edits.

Do transcript spelling edits change recorded speech?

No. Text edits affect transcript output; they do not synthesize or replace audio.

Related tools