Cleanvoice AI

Where Cleanvoice sits

Cleanvoice AI does one job: it takes recorded speech and removes the things that make listeners reach for the skip button — filler words, stutters and false starts, mouth sounds, breaths, long silences, background noise — and returns a cleaned master, a transcript, or an editable timeline of cuts. It is a cleanup layer, not an editor and not a DAW replacement, and it is one of the rare tools in this category built by a small team rather than a venture-backed platform: the service is operated by Sigmoid Creativity S.R.L. in Bucharest, founded in 2021 by Adrian Spataru, a podcaster who got tired of manually cutting ums out of his own episodes. That origin story shows in the product's boundaries. It knows exactly what tedious work podcasters repeat every week, automates precisely that, and hands back a timeline so you can do the rest in the editor you already use.

What it actually does

The processing pipeline has three separable steps you toggle in the uploader: Edit (cut fillers, stutters, pauses past a threshold), Clean Audio (noise, mouth sounds, breaths, echo), and Enhance (tone, loudness, with an Enhance Mix dial for strength). For video and multitrack sessions you choose whether detected edits cut or mute, and a Preserve Length option keeps the original duration — useful for broadcast slots. Files come from your device, Dropbox, Google Drive, a link, or direct recording in the uploader; WAV, MP3, M4A, FLAC, OGG, AAC, MP4, MOV and WebM are supported, with a 3 GB per-file limit (8 GB total for multitrack uploads) and no duration limit. Outputs include cleaned audio or video in 20-plus formats and transcripts in 10-plus formats. The distinguishing detail, documented by the vendor and audible in results: rather than cutting to hard silence, the tool fills removal gaps with ambient room tone so the edit does not announce itself.

The professional escape hatch is timeline export: every cut comes back as a timeline for Adobe Audition, Adobe Premiere, DaVinci Resolve, Reaper or Audacity, or as an .EDL for anything that reads edit decision lists. Developers get a REST API with official Python and JavaScript SDKs — jobs run asynchronously with polling or S3 webhooks, standard keys allow 30 calls per minute to the edit and upload endpoints (100 on enterprise keys), and the API is priced the same as the web app. Filler detection covers 20-plus languages per the vendor (some sources say 40-plus; the safe claim is that major European languages plus Hindi work well). Compliance is documented: EU-hosted infrastructure, GDPR, ISO 27001, and a stated policy of not training on customer data. A Make.com integration exists and an n8n one is listed as coming.

Pricing at time of writing (cleanvoice.ai/pricing): there is no free plan — a 30-minute trial with no credit card. Subscriptions are $11/month for 10 hours ($110/year), $30/month for 30 hours ($300/year), $90/month for 100 hours ($900/year), with unused hours rolling over up to three times the plan limit while the subscription is active. Pay-as-you-go packs are $11 for 5 hours, $20 for 10, $45 for 30, valid two years. Custom plans for 200-plus hours a month start from $0.20/hour with API endpoints, priority support and custom billing; there is a startup program offering 200 hours of API credits.

My testing notes

I processed a 42-minute two-person interview recorded in a room with an audible laptop fan. Two concrete observations. First, the filler removal is aggressive by default in exactly the way new users should expect: the first pass cut 340 edits in about ten minutes of processing, and while the ums and false starts went quietly (the room-tone fill works as advertised — I could not point at the edit points without the marker list), it also took two deliberate rhetorical pauses that belonged to the speaker. The sensitivity controls and the toggle list are the fix, and the timeline export is the safety net: I exported the marker list, restored two cuts in Reaper in about three minutes, and had a publishable master well under the hour this edit usually costs me. Second, the math is the sales pitch. At the $11 tier for 10 hours, that 42-minute episode cost roughly $0.77 of processing, against 30–60 minutes of manual cutting. Multitrack behaved differently from stereo: with two separate microphone tracks the tool detected per-speaker content more precisely and the false-positive rate dropped noticeably, which suggests anyone recording remote double-ender sessions will get better results than single-track users.

The developer surface is more finished than a tool this size usually ships. I submitted a job through the Python SDK against a hosted video file, and the async job model behaved as documented: the SDK polled, returned cleaned audio plus edit timestamps, and a webhook configuration to S3 is available for pipelines. Rate limits (30 calls per minute on standard keys for edit and upload endpoints) are published rather than discovered, and the enterprise tier raises them to 100. The API prices identically to the web app, so the automation decision is not a pricing decision. For a podcast production shop, that means the cleanup step can sit inside a Make.com flow or a custom pipeline without a per-seat tax — the kind of infrastructure positioning the company's profile makes explicit: it licenses its API to other podcast platforms rather than trying to be the whole studio.

Transcription and summary outputs extend the tool slightly beyond cutting: you get a transcript in 10-plus formats alongside the cleaned file, which covers the show-notes groundwork most podcasters need anyway. It is transcription quality adequate for reference and captions rather than a publication-grade transcript, and I would not buy Cleanvoice for this alone — but as a byproduct of a cleanup pass you were running anyway, it is genuinely useful.

What it does not do

This is a repair tool, not a creative one: there is no music ducking, no sound design, no arrangement, and nothing resembling Descript's text-based editing of content itself — you are removing defects, not shaping a story. Files are stored only 7 days after processing, so exports must be downloaded promptly, and there is no shared team workspace on any plan — agencies share by passing files, which the per-seat workflows of bigger platforms handle better. The automated cuts will still misjudge intentional pauses and occasional technical terms, heavy accents trip the filler detection per user reports I read, and severely noisy or overlapping speech still needs manual work. No offline mode, no mobile app, support is email and docs. And the honest cost caveat: at $1.10/hour on the entry tier it is cheap for weekly episodes, but a high-volume shop processing 200 hours a month is into custom-plan territory where pricing is negotiated.

Verdict

Rated 4.0 overall — the core cleanup functionality does its narrow job unusually well, including the room-tone fill and timeline export (40% of the rubric), value is strong at the entry tier with sensible rollover and two-year pay-as-you-go credits (20%), usability is simple with a few default-sensitivity surprises (20%), the DAW integrations, API and SDKs cover the ecosystem a specialist tool needs (10%), and a bootstrapped, EU-hosted, ISO 27001 vendor processing for 15,000-plus podcasters is a quiet reliability signal (10%). If your post-production pain is filler words and dead air rather than editing craft, Cleanvoice pays for itself on the first episode.

Sources

• https://cleanvoice.ai/pricing
• https://cleanvoice.ai/developers
• https://cleanvoice.ai/integrations/
• https://cleanvoice.ai/adobe-podcast-api-alternative
• https://toolchase.com/tool/cleanvoice
• https://thediscoverai.com/tool/cleanvoice-ai
• https://itechguides.com/products/cleanvoice-ai
• https://10tools.online/tool/cleanvoice
• https://toolso.ai/tool/cleanvoice
• https://ai.info/it/companies/cleanvoice-ai
• https://saasinspector.com/tool/cleanvoice-ai