Podcast and Spoken-Word Audio

How to prepare a podcast episode for transcription in REAPER

Published November 6, 2026 · Updated November 6, 2026

Quick answer

First work out which audio your transcript comes from. Apple Podcasts and YouTube transcribe what you publish, so the only preparation that helps there is clearer speech in the final mix. For a tool you upload to, lock the edit, then render a speech-only file from REAPER: music muted, heavy noise reduction bypassed, WAV or FLAC at your project sample rate, one file per person if the tool accepts that. Then give it your names and correct the result by hand. This will not rescue cross-talk, and REAPER does not write the transcript itself.

Transcripts now feed a lot of the work after the edit: show notes, captions, search and automatic chapters. Most advice on getting a better one is about which service to buy. The part you control in REAPER is smaller and comes earlier: what the transcriber actually hears, and what you tell it before it starts.

REAPER’s User Guide describes no speech-to-text feature, so the words come from another tool. REAPER’s job is to hand that tool the right audio, at the right moment in the edit.

Two routes, and only one of them uses your special file

Before you render anything, check where the transcript will be made.

Where the transcript is madeWhat it hearsWhat you can change
Apple Podcasts, automaticThe published episodeThe final mix, and the names in your descriptions. You can replace the transcript with your own.
YouTube, automatic captionsThe audio in the uploaded videoThe final mix. You can review them, or upload your own caption file.
An upload tool (Descript, a speech-to-text service, a local model)Whatever file you give itEverything in this article

Apple says episodes with cross-talk or loud music may be harder to transcribe. So on the automatic routes, the useful fix is in the mix itself: keep the music bed clearly under the voice with ducking in REAPER, and do not talk over each other more than the show needs.

Apple also says to put host and guest names in your show and episode descriptions so they come out spelled correctly. It costs nothing, so do it either way.

Lock the edit first

A transcript with timestamps belongs to one exact version of the episode. Cut a minute out afterwards and every line after that cut points at the wrong moment, so captions and show-note timestamps drift.

Transcribing rough audio can still be useful for finding what to cut. Just treat that as a working draft, and make the transcript you publish from the final length.

Render a speech-only file in REAPER

These steps assume an upload tool. They were checked against the REAPER User Guide; treat the settings as a starting point, because each tool has its own limits.

1

Work in a copy of the project

Use Save project as and add transcript to the name. You are about to mute and bypass things that belong in the real mix, and you do not want them left that way.

2

Mute the music and effects

Mute intro music, beds and stingers, but leave them where they are so no timing changes. Descript’s help says it only transcribes speech and does not support song lyrics, and Apple does not display lyrics either, so music gives the transcriber nothing to do.

3

Bypass heavy noise reduction

Turn off strong denoising on the voice tracks for this render. The next section explains why this goes against instinct, and how to test it on your own show.

4

Set the render

Open File, Render. Use Source Master mix for one file, Bounds Entire project, Channels Mono, format WAV or FLAC, and leave the sample rate at your project rate. Keeping Entire project means the file runs the full length of the episode, so its times match the published version.

Why those formats? Google’s Speech-to-Text guidance recommends a lossless codec, says MP3 and similar lossy formats may reduce accuracy, and asks for 16,000 Hz or higher without re-sampling. AssemblyAI gives similar advice: send audio without converting it, since conversion can lose quality. Those are two services, not a law for every tool, but a normal podcast project at 44.1 or 48 kHz already meets them, so there is nothing to change.

Mono 16-bit WAV at 44.1 kHz runs about 5.3 MB a minute, or a little over 300 MB an hour. If that exceeds your tool’s upload limit, try FLAC first if the tool accepts it, because it stays lossless. Only then use a high-quality MP3, or split the episode. OpenAI’s file transcription, for example, caps files at 25 MB.

The noise reduction trap

Google’s guidance says applying noise reduction before sending audio to its service typically reduces accuracy, and that automatic gain control should not be used. It also says to avoid clipping. That surprises people, because a denoised voice sounds cleaner to us. The likely reason is that denoising leaves artifacts in the speech that a recognizer handles worse than steady background sound.

How far to trust this

That advice is written for Google’s own service, and other engines may behave differently. So test it instead of assuming. Render the same two minutes both ways, with your podcast noise reduction on and off, transcribe both, and count the wrong words. Use whichever wins for your show. We have not run this comparison ourselves.

This applies to the transcription file only. Keep your normal cleanup in the episode people listen to.

Two people: give each voice its own file

If everyone is on their own track, you have a real advantage. Google says that when each person is on a separate channel, you should send each channel separately for the best results. AssemblyAI says the same thing about its multichannel option: it is more accurate than guessing speakers from one mixed channel, because each speaker is processed independently.

Separate tracks are not a cure for bleed between two mics. If each mic clearly hears the other person, expect the same words to turn up on both tracks. Likewise, if two people talked over each other on one mic, no export setting pulls those words apart. Google warns that people talking at once, or at different volumes, may be treated as background noise and ignored.

Tell the tool your names before it starts

Names and show-specific terms are usually the first thing to go wrong. Google says plainly that out-of-vocabulary terms and proper names will not be recognized without hints. Podcaster Matt Birchler reports his co-host’s name being mistranscribed by every model he has tried.

Keep a short list as you edit: people, products, places, and any jargon or spelling your show uses. Then put it wherever your tool accepts it:

Show notes come easier if you also marked the episode as you edited. The markers from adding chapters give you the outline, and REAPER’s Region/Marker Manager has an export option in its right-click menu. The transcript then fills in the detail.

Fix the transcript, then deliver it

Budget for reading it through while the audio plays. Descript says accuracy can reach up to 95 percent with clear audio. That is a best case from a vendor, and even 95 percent leaves about one wrong word in twenty.

DestinationWhat it takesWorth knowing
Apple PodcastsVTT or SRT, linked in your RSS feedSet the episode to display transcripts you provide. Apple says a VTT file lets you identify the speaker on each line. You can download Apple’s version, edit it and submit that.
YouTubeSRT, VTT and othersRelevant if you post the episode as a video. YouTube says to always review automatic captions.
SpotifyVTT, per RSS.comRSS.com’s help says Spotify does not read your feed for this and that you upload in Spotify for Creators. We could not read Spotify’s own page, so check your account.
Your podcast hostVariesCastos, for example, takes .vtt up to 2 MB and refuses SRT. Check yours first.
Watch out

Apple applies quality standards to transcripts you provide, and files that fall short will not be displayed. It asks for no spelling or punctuation errors and for every segment of the episode to be included. So do not skip the read-through, and do not trim the intro or ads out of the text.

What this will not do

If the transcript still comes out wrong

  1. Names are wrong throughout. Add them to the glossary or prompt and run it again, or find and replace.
  2. Lines are missing where people overlap. That is the cross-talk limit. Type those lines by hand from the audio.
  3. Garbled text under the intro. The music was still in. Check your mutes, or lower the bed in the real mix if this is an automatic route.
  4. Timestamps drift later in the episode. The transcript came from an earlier edit, or the render did not use Entire project. Render again and transcribe the final version.
  5. Apple shows no transcript. Apple says to wait at least 24 hours after publishing and to check Podcasts Connect for warnings. Episodes over 10 hours may not be transcribed.
Takeaway

Automatic platform transcripts hear your published mix, so clear speech and quiet music beds are what help there. For an upload tool, lock the edit and render speech only, lossless, with one file per voice if the tool takes it. Test whether your noise reduction helps or hurts, give the tool your names, and read the result against the audio before it goes anywhere.

Sources

  1. Apple Podcasts for Creators, “Transcripts on Apple Podcasts” - automatic transcripts, RSS tag, VTT or SRT, speaker labels, quality standards, cross-talk and loud music, names in descriptions. podcasters.apple.com
  2. Google Cloud, Speech-to-Text best practices (updated October 26, 2026) - sample rate, lossless codecs, noise reduction, AGC, separate channels, overlapping speech, word and phrase hints. docs.cloud.google.com
  3. Cockos, Up and Running: A REAPER User Guide v7.84 (November 2026) - sections 21.3 to 21.6, render source, bounds, stems, formats; Region/Marker Manager. No speech-to-text feature is described there or in the changelog (current release 7.85). dlx.reaper.fm
  4. Descript Help, “Automatic transcription” - accuracy factors, glossary, music and lyrics not supported. help.descript.com
  5. AssemblyAI, FAQ on recommended file type, and on speaker labels versus multichannel. file type, multichannel
  6. OpenAI, “File transcription” - prompts for names and terms, 25 MB file limit. developers.openai.com
  7. YouTube Help, supported subtitle and closed caption files, and automatic captions. formats, automatic captions
  8. RSS.com Help, how transcripts work on Apple Podcasts and Spotify (host documentation, not Spotify’s own). help.rss.com
  9. Castos, Podcasting 2.0 transcripts (October 2026) - .vtt upload rules at that host. castos.com
  10. Matt Birchler, Birchtree post on a podcast transcript fixer (April 2026) - practitioner example of a name being mistranscribed. birchtree.me