Choosing an AI podcast editor is harder than it looks, because most comparisons quote you a transcription price per hour and stop there. That number is the least useful thing about these tools. The figure that actually decides whether a tool works for you is the cost of a finished, publish-ready episode including cleanup, exports and seats.
AI podcast editor comparisons also tend to skip video podcasting entirely, even though a separate video file distributed to YouTube, Spotify and Apple Podcasts is now the default for interview shows. And almost none tell you which speech-recognition engine sits underneath, which is the only way to know whether two accuracy claims describe the same thing. I normalised every price below against one workload: a 60-minute interview episode published monthly, produced by a single editor, needing a transcript, filler-word removal, cleanup and a final export.
Which AI podcast editor should you use? Match the tool to the stage that hurts most
If you want to edit by deleting words, use Descript. If you want the recording itself to be excellent and want video in the same place, use Riverside. If your audio is merely dirty, use Adobe Podcast for free. If your audio is clean but inconsistent in level, use Auphonic, the cheapest serious option at roughly $1.20 per finished hour.
If you need a transcript and nothing else, use Otter.ai for live and short-form work, Rev for legal-grade accuracy, or Sonix for bulk volume at $10 per hour. Skip HappyScribe unless your volume lands neatly on a plan boundary. And note that Podcastle no longer exists under that name: it became Async on 28 January 2026. One caution before you read further.
Not one of the twelve editors below names the speech-recognition model it runs on a public pricing or product page. Every accuracy percentage published for these tools is therefore a vendor claim about an undisclosed model. The only independent word-error-rate benchmarks that exist measure the underlying APIs (Deepgram, ElevenLabs, AssemblyAI, Whisper) and cannot be transferred to the editor you actually subscribe to.
Pricing and accuracy evidence
Prices are from vendor pages checked on 16 October 2026, using annual billing where stated. Cost examples cover software, excluding transcript correction and other editing labour. Regional offers are excluded from the US-dollar comparison.
No test episodes were recorded and no filler-word-removal tests were run for this comparison. Feature descriptions come from vendor documentation; the accuracy section cites an independent speech-to-text benchmark rather than a test of the editors themselves.
The twelve tools, and what each one is actually for
These tools are not interchangeable, and grouping them by what they actually do is the first step to choosing well. Four edit media, four transcribe, two host, one masters audio and one makes clips. Prices are USD on annual billing unless I say otherwise.
Descript: the text editor
Descript is the best AI podcast editor if you think of a transcript as a document you can cut. Deleting a sentence in the transcript deletes the corresponding audio, which turns an hour of waveform scrubbing into an hour of copy-editing.
Its paid plans carry Studio Sound (one-click noise removal and speech enhancement), Remove Filler Words for “um”, “ah” and “like”, Shorten Word Gaps, multitrack transcription and Remove Retakes. Free is $0 with one media hour a month. Hobbyist is $16 per person per month on annual billing with 10 media hours, 400 AI credits, 1080p watermark-free export, 100 GB and one seat. Creator is $24 with 30 media hours, 800 credits, 4K export, 1 TB and up to three seats. Business is $50 with 40 media hours, 1,500 credits, 2 TB, up to five seats and Brand Studio.
The catch is the seat model: Descript prices “per person / month”, so a two-person edit team on Creator pays $48 a month before anyone has published anything. And Descript’s claim that its transcription “delivers up to 95% accuracy” is a vendor claim about an undisclosed model, where “up to” is doing heavy lifting.
Riverside: the recorder that also publishes
Riverside is the best AI podcast editor when recording quality and video distribution are the priority, because it records locally rather than routing audio through a compressed stream. Pro is $29 per month or $24 on annual billing and includes up to 4K desktop video at 44.1 or 48 kHz, no watermark, unlimited recording hours, unlimited transcriptions, text-based editing, AI B-roll, eye contact, silence and filler-word removal, audio enhancement, Magic Clips and show notes, plus publishing to YouTube, Spotify, Apple, Instagram and LinkedIn.
Three limits deserve attention. First, separate audio and video track downloads are capped at 15 hours a month on Pro: a two-person, 60-minute interview consumes two of those, so you get seven such episodes before you are cut off. Second, publishing video to Apple Podcasts requires Grow at $34 per month. Third, an additional editor seat is an add-on, and Riverside states Business is “the only plan with collaborative editing and custom roles built in”. Grow is $39/$34 with two studios and 20 track-download hours; Webinar is $99/$79 with three studios and 25 hours; Business is quote-only, with unlimited studios and tracks, SSO and SOC 2 and ISO 27001 compliance.
Adobe Podcast: the cleanup specialist
Adobe Podcast’s Enhance Speech is the best free AI podcast editor for dirty audio, and its free tier is genuinely usable rather than a demo. The free plan processes audio only, one file at a time, with no strength adjustment, a maximum file duration of 30 minutes (up to 500 MB) and one hour of enhancement per day. Adobe Podcast Studio on free lets you download projects up to 30 minutes, two per day, with podcast-branded audiograms and no download of the originals.
The paid path is Adobe Firefly Pro at $19.99 per month, adding video support for MP4 and MOV, bulk upload, adjustable speech/music/ambience strength, up to four hours of enhanced speech per day with files up to two hours long and 1 GB, unlimited Studio downloads, speaker-separated original-recording downloads and removal of Adobe’s podcast branding.
The 30-minute free file cap is the real constraint: a 60-minute episode must be split and rejoined. On the engine question, Adobe says Studio “transcribes every word using the same industry-leading transcription as Adobe Premiere Pro”, and does not name that model.
Auphonic: the mastering engineer
Auphonic is the best value AI podcast editor for audio quality, and the only tool here priced by a genuinely scalable credit model. Its documentation spells out the audio-processing controls in detail. If you want the wider picture on synthetic-voice tooling, see our round-up of 9 AI voice tools that created professional audio content.
The algorithm list includes an Adaptive Leveler that classifies speech, music and background segments separately so quiet speakers get lifted without pumping up noise; four noise-reduction methods with separately controllable Remove Noise, Remove Reverb and Remove Breaths amounts; Voice AutoEQ with De-Plosive; Bandwidth Extension; and a Studio Voice model that reconstructs codec artifacts and clipped voices, which Auphonic itself flags as beta, warning that “problems might still appear in some recordings”.
Automatic cutting covers silence, fillers (“uhm”, “uh”, “mh”, “ähm”), coughs and foreground music. Critically, an Export Uncut Audio mode writes cut lists for Audacity, Reaper, Adobe Audition, DaVinci Resolve and Final Cut Pro instead of destroying your master. The music cutter is “tuned for music segments longer than about 20 seconds”, so short stings survive.
Pricing in USD on annual billing, which Auphonic says saves 20%: free is 2 hours a month with an Auphonic jingle; S is $11 for 9 hours a month ($1.22 per hour); M is $25 for 21 hours ($1.19); L is $52 for 45 hours ($1.16); XL is $113 for 100 hours ($1.13); XXL is $245 for 250 hours ($0.98). One-time credits run from $12 for 5 hours ($2.40 per hour) down to $171 for 100 hours ($1.71 per hour). Prices exclude VAT. Free users get no multilingual speech recognition, no automatic shownotes, no batch productions, and multitrack limited to 20 minutes.
Otter.ai: a meeting notes app that happens to transcribe
Otter.ai is the cheapest credible subscription transcriber, but it is built for meetings rather than 90-minute interviews. Basic is free with 300 transcription minutes per user per month, but caps a single conversation at 30 minutes and allows only three lifetime file imports. Pro is $8.33 per user per month on annual billing with 1,200 in-app recording minutes and 1,200 imported-file minutes per user, 90 minutes per conversation, unlimited storage and ten monthly imports. Business is $19.99 per user per month with unlimited in-app recordings, 6,000 imported-file minutes and four hours per conversation.
Otter’s own student and teacher discount page settles the monthly list price: it publishes $6.67 per month annual and $13.59 per month monthly as 20% off, implying an annual price near $8.33 and a monthly price of $16.99.
The maths is attractive: $8.33 a month for 20 hours is $0.42 per hour if you use it all, but the conversation ceiling is the catch. A 60-minute show works on Pro; a 90-minute interview format does not.
Rev: transcription and captions, nothing else
Rev is the tool to buy when accuracy is contractual rather than creative, because it also sells human-verified transcription guaranteed at 99%+ and delivered within 12 hours, starting at $1.99 per minute, with legal-formatted transcripts priced by page at $2.25 (rough draft) and $2.50 (premium).
The AI subscriptions are Essentials at $25.49 per seat per month on annual billing ($305.90 a year, three seats included) with 5,000 AI minutes per user per month and English plus Spanish, and Pro at $47.99 per seat per month ($575.88 a year, five seats) with 10,000 minutes and 37+ languages. Free is 45 minutes per month, English only.
Rev runs its own ASR engine, which is independently measurable: “Rev AI” scores 5.9% AA-WER on Artificial Analysis’s leaderboard, the weakest of the engines here on conversational audio. Rev also commits that it does “not sell customer data or allow for customer data to be used to train external LLMs”: note the precise wording, external LLMs.
Sonix: volume transcription
Sonix is the best value AI podcast editor for bulk back-catalogue work at $10 per hour pay-as-you-go, with no seat cost and 30 free minutes to start.
Subscriptions are Core at $25 per month or $275 a year (5 transcription hours plus 5 AI-workspace hours a month, 25 GB, one seat, extra seats at $25 per month); Advanced at $50/$550 (20 hours, 50 GB); and Pro at $80/$880 (40 hours, 100 GB). Additional hours bill at $10 per hour, and translation draws from the same hours.
Sonix’s accuracy claim needs labelling. Its homepage advertises “99% transcription accuracy” and, in the same block, describes unnamed competitors as “85-90% industry average”. No methodology is published for either figure. Sonix does publish the commercially relevant facts: SOC 2 Type II, HIPAA, GDPR, a DPA and a BAA, plus “zero training on your data”.
HappyScribe: cheap entry, brutal volume cliff
HappyScribe is the clearest example of why per-hour transcription prices are useless. Its plans bill transcription, subtitling and translation as an annual allowance, so your effective cost per hour depends entirely on where your volume lands.
Basic is $8.50 per month billed annually, giving 1,440 minutes a year (24 hours, or two a month) plus one seat and watermark removal. Pro is $19 per month, giving 7,200 minutes a year, or ten hours a month, plus three seats. Business is $59 per month, giving 1,200 hours, plus five seats.
Run the numbers. Basic gives $4.25 per finished hour at two episodes a month. Pro gives $4.75 at four a month: twice Basic’s per-episode cost, because you have outgrown it. Pro only becomes efficient at its full 10 hours a month, where it costs $1.90. Business at full volume is $0.59. The free tier is not a viable podcast plan at all: 45 minutes per meeting recording, a 10-minute free trial, and watermarked MP4 exports.
Async: formerly Podcastle, now credit-based
Podcastle rebranded to Async on 28 January 2026, announced by the company with $23.5M raised to date from AI Fund, Mosaic Ventures, RTP Global, Point Nine and Sierra Ventures. The vendor states that “all accounts, projects, pricing tiers, and billing structures remain unchanged”, but the scope widened well beyond podcasting into general AI video, voice and developer workflows. Any comparison still recommending Podcastle is out of date.
On annual billing, advertised as saving 40%: Essentials is $11.99 per month with 450 AI credits, 4 hours of audio and video recording, 20 GB storage, Magic Dust enhancement, noise reduction, silence removal and English-only subtitles. Pro is $23.99 with 1,200 credits, 20 recording hours and 120 GB. Teams is $49.99 per creator with 3,000 credits, 50 recording hours, 1 TB, dubbing and lipsync and brand kits. Business is quote-only.
The allowances that actually bite are separate from the credits: transcription caps at 10 hours on Essentials, 25 on Pro and 100 on Teams, and remote multitrack recording at 2, 20 and 50 hours. Export is up to 720p on free, full HD on paid and 4K on Pro, with audio at 320 kbps MP3 or 1411 kbps WAV on paid against 160 kbps MP3 on free, and free is limited to a single exported file.
Buzzsprout and Transistor: hosts that include transcripts
If you only need a transcript to publish, the cheapest route is to let your host do it. Buzzsprout states that it “automatically creates transcripts for new episodes on all current paid plans at no additional cost”, includes them in the RSS feed and on the website. Audio is $15 per month on annual billing for up to 72 upload hours a year; Audio + Video is $25 with 1080p video to Apple Podcasts, Spotify and YouTube; Multi-Podcast is $30 for five or ten shows with 180 hours a year. Free gives two upload hours a month and 90 days of hosting. Both hosts include unlimited team members, which no per-seat editor does.
Transistor is $19 per month on annual billing for Starter (20,000 monthly downloads), $49 for Professional (100,000, plus auto-post to YouTube) and $99 for Business (250,000, plus branding removal), with unlimited podcasts, no storage limits and no episode cap. Transcription is an add-on: 5 hours a month for $5, or 20 hours for $20: $1.00 per hour either way. Buzzsprout’s Cohost AI and Magic Mastering are paid add-ons whose prices are not published on its pricing page.
Podsqueeze: the clip factory
Podsqueeze is not an editor; it is a repurposing tool that happens to transcribe. Its value is turning chapters into vertical clips, audiograms and social posts, plus hosted landing pages.
Starter is $8.99 per month with 120 minutes a month (which roll over) and eight clips; Pro is $49 with 320 minutes and 20 clips; Agency Lite is $89 with 600 minutes and 40 clips. The page also displays a “30% OFF” annual toggle, so which figure applies to annual billing is ambiguous from the page itself.
For a four-episode-a-month show you exhaust Starter’s 120 minutes and land on Pro at $49, or $12.25 per finished hour. It is a clip tool with a transcription bill attached. For the wider repurposing question, see our guide to 10 AI video generation methods for viral marketing content.
Which speech-recognition engine runs under the hood?
The table below compares speech-to-text engines using Artificial Analysis’s Speech-to-Text leaderboard on 16 October 2026.
AA-WER is the percentage of words transcribed incorrectly: lower is better. The v2 index blends three datasets: AA-AgentTalk at 50%, VoxPopuli-Cleaned-AA at 25% and Earnings22-Cleaned-AA at 25%. Prices are per 1,000 minutes of audio.
| Engine (non-streaming) | AA-WER v2 | Price USD / 1,000 min | Equivalent per hour |
|---|---|---|---|
| ElevenLabs Scribe v2 | 2.2% | $3.67 | $0.22 |
| Zoom Scribe V2 Pro | 2.8% | $3.33 | $0.20 |
| AssemblyAI Universal-3 Pro | 3.1% | $3.50 | $0.21 |
| AssemblyAI Universal | 3.8% | $2.50 | $0.15 |
| Soniox v5 | 3.8% | $1.66 | $0.10 |
| OpenAI Whisper Large v2 | 4.1% | $6.00 | $0.36 |
| Deepgram Nova-3 | 5.2% | $4.30 | $0.26 |
| Rev AI | 5.9% | $3.33 | $0.20 |
I cross-verified four of these against each vendor’s own pricing page and they match exactly: Deepgram Nova-3 at $4.30 per 1,000 minutes equals Deepgram’s published $0.0043 per minute; ElevenLabs Scribe v2 at $3.67 equals $0.22 per hour; AssemblyAI Universal at $2.50 equals $0.15 per hour; Universal-3 Pro at $3.50 equals $0.21 per hour.
Three caveats matter. This measures APIs, not editors. The datasets are not podcast audio: agent conversations, multilingual parliament speech and earnings calls. And none of the consumer editors below is in this table, because none discloses its engine.
| Editor | Disclosed ASR engine |
|---|---|
| Descript | Not disclosed. Claims “up to 95% accuracy”. |
| Riverside | Not disclosed |
| Adobe Podcast / Firefly Pro | Not disclosed. Says only “the same industry-leading transcription as Adobe Premiere Pro”. |
| Auphonic | Not disclosed. Homepage testimonial from TU Graz describes a Whisper integration; Auphonic publishes no model list. |
| Otter.ai | Not disclosed |
| Rev | Rev AI, its own engine, independently benchmarked at 5.9% |
| Sonix | Not disclosed. Claims “99% accuracy” with no published methodology. |
| HappyScribe | Not disclosed |
| Async | Not disclosed |
| Buzzsprout | Not disclosed |
| Transistor | Not disclosed |
| Podsqueeze | Not disclosed |
The only tool you can run entirely yourself is OpenAI’s Whisper, whose code and weights are MIT licensed. whisper.cpp is a C/C++ port that runs fully offline on a Raspberry Pi, an iPhone or CPU-only hardware, with the large model needing about 3.9 GB of memory and the base model about 388 MB. That is the privacy-respecting option, and it is free.
What a finished episode really costs
Here is the table the industry is missing. Prices are USD, annual billing unless noted, for a 60-minute episode. “Finished hour” means transcript plus filler removal plus cleanup plus an export you can actually publish.
| Tool | Plan used | Monthly cost | Included allowance | $/finished hr at 4 eps/mo | $/finished hr at 12 eps/mo |
|---|---|---|---|---|---|
| Descript | Hobbyist | $16 | 10 media hrs/mo, 1 seat | $4.00 | $1.60 |
| Descript | Creator | $24 | 30 media hrs/mo, 3 seats | $6.00 | $2.00 |
| Riverside | Pro | $24 | Unlimited recording + transcription | $6.00 | $2.00 |
| Riverside | Grow (Apple video) | $34 | Unlimited + 20 track hrs | $8.50 | $2.83 |
| Async | Essentials | $11.99 | 10 transcript hrs, 4 recording hrs | $3.00 | Needs Pro |
| Async | Pro | $23.99 | 25 transcript hrs, 20 recording hrs | $6.00 | $2.00 |
| Adobe Podcast | Free | $0 | 1 hr/day, 30-min files | $0.00 (files must be split) | $0.00 |
| Adobe Firefly Pro | Pro | $19.99 | 4 hrs/day, 2-hr files | $5.00 | $1.67 |
| Auphonic | S | $11 | 9 credit hrs/mo | $2.75 | $1.22 |
| Auphonic | M | $25 | 21 credit hrs/mo | $6.25 | $2.08 |
| Otter.ai | Pro | $8.33 | 20 hrs, 90-min ceiling | $2.08 | $0.69 |
| Rev | Essentials | $25.49 | 83.3 hrs, 3 seats | $6.37 | $2.12 |
| Sonix | Core | $25 | 5 hrs/mo | $6.25 | Needs Advanced |
| HappyScribe | Pro | $19 | 10 hrs/mo, 3 seats | $4.75 | $1.90 |
| Buzzsprout (host) | Audio | $15 | 72 upload hrs/yr | $1.25 | Hosting is not per-episode |
| Transistor (host) + TX add-on | Starter + 5 hrs | $24 | Unlimited hosting + 5 hrs TX | $2.00 | $2.00 |
The pattern is awkward. The cheapest software is often the tools that do the least. Adobe’s free tier costs nothing and cannot process a 60-minute episode without splitting it. Otter at $2.08 per finished hour is cheaper than Descript and cannot cut your video.
The hidden costs nobody lists
The headline price is roughly half the real one, because seats, watermarks, per-file ceilings and non-rolling credits all sit outside it. These are the five that catch people out.
- Per-seat multiplication. Descript prices per person; Sonix charges $25 more per extra seat; Rev is per seat; Async Teams is $49.99 per creator. A two-person team roughly doubles the Descript and Otter numbers above.
- Watermarks in free tiers. Descript free exports 720p with a watermark; HappyScribe free exports watermarked MP4; Async free allows one export file at 160 kbps MP3.
- Caps that are not about transcription. Adobe free: 30-minute files, one hour a day. Otter Pro: 90 minutes per conversation. Riverside: 15 separate-track hours a month. These are the limits that actually break a show.
- Credits that reset. Otter’s transcription minutes, Buzzsprout’s upload hours and Auphonic’s credits all expire at the billing boundary rather than accumulating. Podsqueeze’s minutes do roll over: one of the few that do.
- Unpublished add-on prices. Buzzsprout lists Cohost AI and Magic Mastering without prices on its public page.
How to pick: a seven-step decision path
Choose by elimination, not by feature scorecard: rule out tools your episode length or format disqualifies, then price what is left. The steps are in the order that removes the most options fastest.
- Fix the episode length and frequency first. Otter Pro caps a single conversation at 90 minutes and Basic at 30, so a 60-minute show already disqualifies the free tier. Adobe’s free file cap of 30 minutes forces a split.
- Decide whether you publish video. If yes, Riverside Grow ($34/month) or Buzzsprout Audio + Video ($25/month) are the two paths that actually distribute video to Apple Podcasts, Spotify and YouTube.
- Count your editors, not your hosts. One editor on Descript Hobbyist is $16; two on Creator is $48. Riverside charges extra editor seats as an add-on and reserves collaborative editing for Business.
- Budget for cleanup and mastering, or accept you will not have it. Otter, Rev, Sonix and HappyScribe transcribe. They do not level, denoise or normalise loudness. Auphonic’s S plan at $11 a month is the cheapest credible way to get all three.
- Decide on voice regeneration, and read the terms if you say yes. Under Descript’s terms a “Consenting Speaker” must have given permission for their voice to train an AI Voice and be synthesised, and Descript’s “employees, vendors, and contractors (including Amazon Mechanical Turk workers) may listen to audio samples of your Training Audio”. That is a materially different privacy posture from transcription.
- Check whether your data will train someone’s model. Otter says yes by default. Riverside says yes unless you are a Business customer with a written opt-out. Descript says yes unless you opt out. Sonix says no.
- Run a two-week pilot on real material, not a demo file. Use an episode with genuine crosstalk, two remote guests on poor connections, and a commercial intro track. Compare the exports by ear on your audio.
What breaks: failure modes and gotchas
Filler-word removal deletes real speech. This is the most common complaint with text-based editing. When two people talk over each other, deleting a word from one speaker’s line can take a chunk of the other with it. Auphonic sidesteps the worst of it by recommending filler cutting together with its Dynamic Denoiser or Speech Isolation to avoid cut artefacts, and by offering Export Uncut Audio as a non-destructive alternative.
Studio-sound restoration is lossy and the vendors admit it. Auphonic labels Studio Voice beta and warns that “problems might still appear in some recordings”. Adobe’s framing is that it makes bad audio “sound as if it were recorded in a professional podcasting studio”: a simulation, not a recovery.
Denoising cannot undo the room. Auphonic’s usage notes are blunt: keep the microphone close, do not level or gate before denoising, and be aware that “high noise-reduction amounts might result in artifacts”. Built-in laptop microphones in a hard room are beyond what any model fixes.
Loudness targets are not universal, so a tool that hides the target is a problem. Auphonic publishes targets spanning Netflix at -27 LUFS, EBU R128 at -23, ATSC A/85 at -24, Spotify and Tidal at -14, mobile at -16, and RMS targets of -18 to -23 dB RMS for Audible and ACX. Check your distribution target before committing.
Cut-list export is fiddlier than it looks. Auphonic warns cut lists “will always be relative to your input files” and that DaVinci Resolve needs the media imported to the media pool first. Descript caps web-link export at one hour on Free and Hobbyist and three hours on Creator and Business: irrelevant if you self-host, fatal if you planned to share a link.
Consent, data protection and disclosure
The following summarises vendor policies and published rules. Confirm the requirements for your own production.
Your guests are data subjects, and Otter says so itself. Its privacy policy instructs users directly: “If you provide an Audio Recording, this may contain the Personal Information of third parties. Before you do so, please make sure you have the necessary permissions from your co-workers, friends or other third parties before sharing Personal Information or referring them to us.” Riverside’s policy carries a dedicated “Notice of Collection of Voice and Face Data” section that expressly engages the CCPA’s sensitive personal information framework.
Training on your audio is standard unless you stop it. Otter processes data to “improve and monitor the Services, including training our proprietary AI technology on de-identified audio recordings and on transcriptions (which may contain Personal Information)”, and uses “Data labeling service providers who provide annotation services and use the data we share to create training and evaluation data”. Its cloud compute and storage run on Amazon Web Services in the United States, it relies on Standard Contractual Clauses for transfers outside the EEA and UK, and it is certified under the EU-U.S., UK Extension and Swiss-U.S. Data Privacy Frameworks.
Riverside says “Content Data may be used by Riverside for improving the Services, including for training Riverside’s Artificial Intelligence and Machine Learning models”, but adds that Content Data “does not include personal data contained in content of Customers on a business plan who have opted out of AI training pursuant to a written agreement”. Descript’s terms grant it the right to use your input and output “to train, improve, and develop its models and technologies (unless you have opted out of such use)”, while stating its agreements stop third-party providers training their own models. Sonix advertises “zero training on your data”.
The EU has now ruled on voice and AI models. The EDPB adopted Opinion 28/2024 on 17 December 2024, requested by the Irish Data Protection Commission. It holds that AI models trained with personal data “cannot, in all cases, be considered anonymous”, and gives the example of “a generative model fine-tuned on the voice recordings of an individual to mimic their voice”.
The EU AI Act’s transparency rules now apply. The European Commission states the AI Act entered into force on 1 August 2024 and became applicable on 2 August 2026, with transparency rules taking effect that month and the AI Office taking enforcement powers from the same date. Article 50(2) requires providers of systems generating synthetic audio, image, video or text to mark outputs in a machine-readable format as artificially generated. Article 50(4) requires deployers producing deep fakes to disclose that fact. There is an exception where a system “perform[s] an assistive function for standard editing or do[es] not substantially alter the input data provided by the deployer”.
My reading (flagged as inference, not guidance) is that trimming a filler word falls inside “standard editing”, while regenerating a host’s line with a cloned voice is generating synthetic audio. The Commission’s voluntary Code of Practice on Marking and Labelling of AI-generated Content is the practical reference.
California has moved too, but probably not for you. The California Privacy Protection Agency announced on 23 September 2025 that the Office of Administrative Law had approved regulations covering cybersecurity audits, risk assessments, automated decisionmaking technology (ADMT), insurance and CCPA updates. They took effect 1 January 2026, with risk-assessment compliance beginning then and attestations due to the CPPA by 1 April 2028, and ADMT requirements for significant decisions beginning 1 January 2027. A podcast transcription pipeline is not making significant decisions about consumers, so ADMT almost certainly does not reach you. Notice and risk-assessment duties are more plausibly in play.
Music and second-party rights are where podcasters actually get taken down. Bundled libraries are still licences: Descript Creator includes “unlimited access to royalty-free stock media library” and Adobe Podcast offers “pre-edited royalty free music… podcast-ready sound collections”, but scope differs by vendor and plan. Read the library’s terms, not the feature list. And if you run aggressive music removal over a master with a licensed intro under it, Auphonic’s documentation says its cutter is tuned for segments longer than about 20 seconds: a 6-second stinger survives, a 40-second outro does not. On hosting, Transistor’s terms make users “solely liable for any content they upload”, while its pricing FAQ states you retain ownership.
The bottom line
For a weekly interview show with video, I would record in Riverside Grow at $34 a month for Apple Podcasts video distribution and 20 separate-track hours, and add Auphonic S at $11 for mastering and cut lists. That is $45 a month, or $11.25 per finished hour including hosting, with a transcript included.
If the show is audio-only and I edit by text, Descript Creator at $24 a month with three seats is the pragmatic choice, understanding that “up to 95% accuracy” is a claim about a model nobody will name.
If budget is binding and I have technical patience, Auphonic plus an ElevenLabs Scribe v2 (see our ElevenLabs review for the wider platform) or AssemblyAI pipeline runs at roughly $1.40 per finished hour, and running Whisper locally removes the data-protection question entirely at no licence cost.
What I would not do is buy a transcription service and expect an editor. Otter, Rev, Sonix and HappyScribe produce text. They do not clean audio, they do not cut video, and at these prices their subscriptions cost several times what a metered API does.
FAQ
The ten questions readers ask most, answered from the same pricing and policy pages cited above.
Which AI podcast editor is the best all-rounder in 2026?
Descript at $16 per person per month is the strongest text-based editor. Riverside at $24 is strongest if recording quality and simultaneous video distribution matter, because it records locally in up to 4K at 48 kHz. Adobe Podcast’s free tier wins on pure audio cleanup at one processed hour a day. Auphonic from $11 a month is the cheapest serious mastering option at roughly $1.20 per finished hour.
How much does it cost to produce one finished hour of podcast audio?
For a monthly show, Descript Hobbyist at $16 and Auphonic S at $11 each work out to roughly $16 and $11 per finished hour because you use only a small slice of the allowance. Production ASR APIs are far cheaper: Deepgram Nova-3 at $0.0043 per minute is $0.26 per hour, and AssemblyAI Universal at $0.15 per hour is cheaper still.
Which podcast editor removes filler words and silence best?
Descript, Auphonic and Riverside all ship automatic filler and silence removal. Auphonic is the most configurable, with separate algorithms for silences, fillers, coughs and music, plus an Export Uncut Audio mode that hands you cut lists for Audacity, Reaper, Audition and DaVinci Resolve instead of destroying your master. Descript’s Remove Filler Words acts on the transcript, which is faster but makes every cut a hard delete.
What AI model does Descript, Otter or Riverside actually use for transcription?
None of them says. None of the twelve editors in this comparison names a third-party speech-recognition model on its public pricing or product page. Rev is the closest to an exception because its Rev AI engine appears under its own name on Artificial Analysis’s leaderboard at 5.9% AA-WER. You therefore cannot transfer a benchmark score to an editor’s product.
Is there an independent accuracy benchmark for podcast transcription?
For speech-to-text APIs, yes. Artificial Analysis publishes AA-WER v2, blending AA-AgentTalk (50%), VoxPopuli-Cleaned-AA (25%) and Earnings22-Cleaned-AA (25%). On 16 October 2026 it placed ElevenLabs Scribe v2 at 2.2%, AssemblyAI Universal-3 Pro at 3.1%, Universal at 3.8%, OpenAI Whisper Large v2 at 4.1%, Deepgram Nova-3 at 5.2% and Rev AI at 5.9%. For consumer podcast editors there is no equivalent benchmark.
Can I make a video podcast entirely inside one of these tools?
Yes, with three caveats. Riverside Pro records up to 4K at 44.1 or 48 kHz with no watermark, but Apple Podcasts video distribution requires Grow at $34 per month. Descript’s paid tiers export 1080p on Hobbyist and 4K on Creator, but the free plan watermarks 720p. Async offers 4K on Pro, but its free tier allows one export file at 160 kbps MP3.
Do these tools train their AI models on my podcast audio?
Some do, by default. Otter’s privacy policy states it trains its AI on de-identified audio and on transcriptions that may contain personal information. Riverside says Content Data may be used to train its models, with an opt-out for Business customers under written agreement. Descript’s terms grant it the right to train unless you opt out. Sonix advertises zero training on customer data.
What are the hidden costs that podcast-editor comparisons usually miss?
Four, mainly. Per-seat pricing multiplies everything: Descript is per person, Sonix adds $25 per extra seat, Rev is per seat, Async Teams is $49.99 per creator. Watermarks and export ceilings hide in free tiers. Non-transcription caps bite hardest: Otter Pro limits a conversation to 90 minutes, Riverside allows 15 hours of separate track downloads a month. And credits rarely roll over.
Do I need a standalone transcription tool if my editor already transcribes?
Only if you need something the editor does not do. Riverside and Descript include unlimited transcriptions, and Buzzsprout generates transcripts free on every paid plan and puts them in the RSS feed. Buy a separate service when you need verbatim accuracy for legal or research use, when you need a format your editor cannot export, or when a metered API at $0.15 to $0.26 per hour is simply cheaper than a seat licence.
What is the cheapest reliable way to get publish-ready podcast audio?
A two-tool stack. ElevenLabs Scribe v2 at $0.22 per hour or AssemblyAI Universal at $0.15 per hour handles the text; Auphonic at $1.22 per finished hour on its $11 S plan handles levelling, denoising and loudness normalisation to a target such as -16 LUFS. That lands near $1.40 per finished hour before your editor’s time, versus $16 for Descript Hobbyist on a monthly show.
Sources
- Descript Pricing: plan prices, media hours, AI credits, export resolutions, storage, seats
- Descript: AI Transcription: plan detail, 25 transcription languages, “up to 95% accuracy” vendor claim
- Descript Security: SOC 2 alignment, GDPR/CCPA, AES-256, TLS 1.2, opt-in AI features
- Descript Terms of Service: Input/Output ownership, training clause, Consenting Speaker and Training Audio terms
- Riverside Pricing: Pro/Grow/Webinar/Business pricing, track-download hours, recording specs, AI credits
- Riverside Privacy Policy: Content Data training clause, Voice and Face Data notice, Business opt-out
- Riverside is now SOC 2 Type 2 Compliant: SOC 2 Type 2 status
- Adobe Podcast Plans: free vs Firefly Pro limits, $19.99 price, duration and file-size caps
- Adobe Podcast Enhance Speech: Firefly Pro feature list, 4-hour daily cap
- Adobe Podcast home: Studio transcription claim, 16-bit 48k WAV multitrack, royalty-free music library
- Auphonic Pricing: credit tiers, one-time credits, feature comparison, VAT note
- Auphonic: Audio Post Production Algorithms: Adaptive Leveler, denoisers, Studio Voice beta, automatic cutting, cut-list formats, LUFS targets
- Auphonic: Loudness Specifications: loudness target examples
- Otter.ai Pricing: plan limits, conversation caps, workspace sizes, education discount figures
- Otter.ai Privacy Policy: training on de-identified audio, data labelling providers, DPF certification, AWS/US transfer, guest permission warning
- Rev Pricing: seat pricing, AI minutes, human transcription rates, “no training on external LLMs”
- Sonix Pricing: pay-as-you-go, tier allowances, extra seats, extra-hour rate, compliance badges
- Sonix home: “99% transcription accuracy” vendor claim, zero training on customer data
- HappyScribe Pricing and Rates: annual minute allowances, seat counts, export formats, free-tier watermarking
- Async Pricing: annual plan prices, credits, recording hours, storage
- Async: Podcast transcription: transcription feature page (no engine named)
- Podcastle rebrands as Async: 28 January 2026 rebrand, funding, unchanged pricing tiers
- Buzzsprout Pricing: plan prices, upload hours, free transcripts, add-on disclosure, rollover policy
- Transistor Pricing: plan prices, download limits, transcription add-on rates, content ownership statement
- Transistor Terms of Service: user content liability, DPA reference
- Podsqueeze Pricing: plan prices, minutes, clip allowances, rollover
- Podsqueeze home: audio enhancement, clip generation, landing pages
- Deepgram Pricing: Nova-3 and Whisper per-minute pricing, diarization and keyterm add-ons, compliance
- AssemblyAI Pricing: Universal and Universal-3.5 Pro hourly rates
- ElevenLabs API Pricing: Scribe v2 and Realtime hourly rates, keyterm and transcript editing add-ons
- ElevenLabs Transcription docs: Scribe output format with word timestamps and speaker IDs
- Artificial Analysis: Speech-to-Text Leaderboard: AA-WER v2 methodology, engine scores and per-1,000-minute prices
- OpenAI Whisper repository: MIT licence, model sizes, language coverage
- whisper.cpp repository: offline inference, platform support, memory requirements
- California Privacy Protection Agency: California Finalizes Regulations: 1 January 2026 effective date, risk-assessment and ADMT timelines
- EDPB Opinion 28/2024 on AI models: anonymity test, voice-recordings example, legitimate interest test
- European Commission: AI Act regulatory framework: applicability from 2 August 2026, transparency rules, AI Omnibus timeline
- EU AI Act Article 50: synthetic content marking and deepfake disclosure obligations and standard-editing exception
