Purplelink
← All apps
Tapefolio app icon
macOS 26 · Apple silicon Free 7-day trial · $29.99 to keep

Tapefolio

Tapefolio transcribes interview and meeting recordings on your Mac, labels the speakers, helps you replace names and other identifiers, and reads scanned pages and photos into searchable text. Your files are not uploaded.

Try it free for 7 days

Checkout shows the price in your local currency, and you can switch back to US dollars.

Add any second Mac Suite app and save 25% on both, or any third for 40% off all three. Choose the apps.

The trial is the complete app for seven days from the first time you open it, with no account and nothing to enter. Buying is a single $29.99 payment, with every update included. You get a license key by email and on the page after you pay. Paste it into the same app and it stays valid, with no second download. One key works on two of your Macs. Not useful? Email ben@purplelink.llc within 14 days of purchase (terms). Version 1.6.0, released November 18, 2026. 44 MB disk image. Needs macOS 26 or later on a Mac with Apple silicon.

Comparing options? See Mac apps for transcription, text to speech and OCR compared.

Also in the Mac Suite: ModernTex, Outbound Veil, Legroom, Keyfeel, Tapefolio, Panely, Summonly, Née, Lenses and Vitae Plus for life, $59.99 once.

Tapefolio's review window on an interview transcript. The left panel lists two speakers with their share of the words and seven identifiers found, none reviewed yet. The transcript fills the right, with names, a street and a first name underlined, and an audio player along the bottom.
The review screen on a demo recording: an invented interview read by two computer voices. The draft is not perfect, and the first pass did not mark every identifier.

Forty-five seconds in the app

The review screen on a demo interview, replacing or keeping each identifier, the export sheet, Batch OCR and the model settings. The footage is the running app. The interview is invented and read by computer voices.

How it works

You add a recording

Open audio files, or record from the microphone inside the app. Tapefolio transcribes on your Mac and keeps the time of every word.

You check it against the audio

The review screen plays the recording beside the text and highlights the word being spoken. You correct words, name the speakers and merge two labels that are one person.

You decide what is replaced

Identifier removal marks what it found. You replace or keep each one, or replace everything left, and then export. The key that maps the codes back is saved in a separate file.

What it looks like

Captures of the running app. The recording in the first four is an invented interview read by two computer voices. The letters in the OCR capture are generated test files. Every name and number is made up. The draft has transcription errors, and the first pass missed some identifiers.

The review window before any decisions. Speakers are named Speaker 1 and Speaker 2, with 46% and 54% of the words. Seven identifiers are listed on the left with their times and codes, and the progress line reads 0 to replace, 0 left in, 7 not reviewed.
The review screen before any decisions. Seven identifiers are found and none is reviewed. The draft has errors: the clinic line reads “Brightwater Clinic Indecator”, and in this first pass the app did not mark “Decatur” (transcribed there as “Indecator”) or “Brightwater Clinic” as identifiers.
The same window after decisions. The speakers are renamed Interviewer and Participant. Five identifiers have a filled check mark, Elena is struck through and marked left in, and the progress line reads 5 to replace, 1 left in, 1 not reviewed. A Check these too list at the bottom offers Brightwater and Clinic.
After deciding case by case: five identifiers to replace, one left in (Elena) and one not yet reviewed. The speakers have been renamed. A “Check these too” list at the bottom of the panel offers more words, here “Brightwater” and “Clinic”, that you can remove as an identifier type. Anything you do not review is replaced when you export.
The transcript with the replacements shown. The text reads Dr. [PERSON 1], [PERSON 2], [ADDRESS 1], [PHONE 1], [EMAIL 1] and [PERSON 3] in place of the names, street, phone number and email address. Elena and the words Brightwater Clinic Indecator are still readable.
The preview with codes in place of the identifiers marked to replace. Elena was kept, and “Brightwater Clinic Indecator” is still readable because it was never marked. Detection misses some identifiers, so you read the result.
The Export sheet over the review window. Rows for Format (Web page, HTML), Look (Paper, light, serif), Title, Show timestamps, Add the draft notice at the top, Use MAXQDA time markers and Replace identifiers with codes. A note says 5 replaced, 1 left in, 1 not yet reviewed, and that the key that undoes the replacement is saved next to the export.
The export sheet. It offers a format and a look, timestamps, a draft notice, MAXQDA time markers and replacing identifiers with codes. It states how many were replaced, left in and not yet reviewed, and that the key that undoes the replacement is saved next to the export, to be kept apart from it.
The Batch OCR window. A left panel sets the workflow, what to read, how to read it and which outputs to make. In the middle a list of nine files shows each file's page count and confidence, with 97% for the receipt and 100% for the others. A page image and the recognized lines of a letter sit on the right.
Batch OCR on nine generated test files, 11 pages. The panel on the left sets the outputs, the list shows each file's confidence (97% for the receipt, 100% for the rest), and the recognized lines of the selected letter are on the right. The names and numbers in the letter are invented.
Settings, Models. Speech to text is set to Apple, built into macOS, many languages, and speaker separation to Standard. Under Models on this Mac, Whisper base English offers a 150 MB download, and NVIDIA Parakeet English and Whisper large v3 turbo are installed with a Delete button each.
Settings, Models. Speech to text is Apple’s recognition until you choose another. Whisper base is a 150 MB download here, and Parakeet and Whisper large v3 turbo are installed on this Mac, each with a Delete button.

What it does

TranscriptionTranscribes audio and video files on your Mac, such as .m4a, .mp3, .mp4 and .mov, with word timing. Until you download a speech model it uses Apple's on-device speech recognition, and macOS may fetch Apple's own model once the first time. Better models are optional downloads in Settings, from Hugging Face, each once: NVIDIA Parakeet, Whisper base and Whisper large-v3 turbo. Settings can delete any of them. Each interview can have its own spoken language. Whisper large-v3 turbo can detect it, and Apple's speech recognition uses the language you pick. Parakeet and Whisper base are English only.
SpeakersLabels who is speaking, using the standard speaker-separation models that come with the app. In the review screen you name each speaker once and merge labels that belong to one person. NVIDIA Nemotron 3 Diarization is recommended instead. It was far better on the real two-person recordings measured below, Tapefolio uses it automatically once it is installed, and it is a 190 MB download.
Speaker memoryOptional, and opt-in for each project. Tapefolio requires a recorded consent note before it saves a voice. Voices are stored encrypted on your Mac, the app only suggests matches for you to confirm, and you can delete them.
Identifier removalFinds names, places, organizations, emails, phone numbers, addresses, links, ID numbers and dates of birth, and replaces them with consistent codes. You review case by case: replace or keep each one, or replace all that remain. The key file that maps the codes back is kept separately. It misses some, so you read the result.
Review screenAn audio player next to the transcript, with the current word highlighted. Turns are editable, and you can rename or merge speakers and go through the identifier review without leaving the screen.
ExportWord, HTML, Markdown, plain text, SRT, VTT and JSON, and files made to import into MAXQDA, NVivo, ATLAS.ti and Dedoose, following each program's own import instructions.
RecordingRecord from the microphone in the app, then transcribe straight into the review screen.
Call recordingRecords a Zoom, Microsoft Teams, Google Meet or FaceTime call with the call's sound on its own track beside your microphone, and no bot joins the call. Your words are labelled You from the microphone track, and the people on the call are told apart on the call track. You start and stop it yourself, and before each call Tapefolio asks you to confirm that everyone has agreed to be recorded. Recording laws differ, and in some places everyone on the call must agree.
Find and word listFind any word or phrase in the open transcript and jump the audio to it. A word list of names, places and jargon helps Whisper and Apple's recognition spell them, and fixes words it hears wrong, such as a participant's pseudonym.
DictationPress Control-Option-Space in any app, speak, and press it again. The text is transcribed on your Mac and typed where the cursor is, with your dictionary, snippets and a choice of styles. The audio is deleted once it is transcribed. Typing for you needs Accessibility permission; without it the text goes to the clipboard.
Draft summaryOptional. Apple's on-device model writes a short overview and a list of topics with times. A quote is shown only when it matches the transcript word for word. It is a draft that can be wrong or leave things out. It needs a Mac with Apple Intelligence turned on, and nothing is downloaded or sent anywhere.
AI appsOptional. Claude Desktop and other apps that support MCP can search and read your saved transcripts, read only. By default names and other identifiers are replaced by codes such as [PERSON 1], speakers show as Speaker 1 and so on, and file names are never shown. Key files are never read. What the AI app reads is sent to its own servers, for example Anthropic's for Claude, so check a transcript before you share it this way. It is off until you add it to the AI app.
Batch OCRReads photos, scans and PDFs, one file or a whole folder, and keeps two-column pages in reading order. Output is a searchable PDF with an invisible text layer, or text, Markdown, Word, JSON or CSV. You can save a workflow and run it again, and replace identifiers in the text with a separate key file. Each run writes a provenance record with a SHA-256 for every input file.

What it will not do

Get every word rightIt will mishear words and mislabel speakers. Overlapping voices, quiet speech and short replies such as okay and yep are where it loses the most. Check any text you plan to quote against the recording.
Anonymize a transcriptIdentifier removal finds many identifiers and misses some, and it cannot tell when a detail identifies someone only in context. It does not make a study compliant. That is for you and your ethics board. Read every transcript before you share it.
Read handwriting or phone photos reliablyOCR has been measured only on generated test files. Real handwriting and real phone photos of pages have not been tested.
Promise results in other languagesWhat has been measured is English meeting audio. Whisper large-v3 turbo and Apple's speech recognition transcribe other languages, but accuracy has been measured in English only. Accents and a quiet two-person interview have not been measured.
Run on an Intel Mac or an older macOSIt needs macOS 26 or later and Apple silicon. There is no Windows, iPhone or iPad version.
Send your files anywhereThere is no cloud transcription and there are no accounts. Nothing you open in Tapefolio is uploaded. The one exception is your choice: if you connect an AI app such as Claude Desktop, the text it reads goes to that app's servers.

How well it works

These are small checks, run on an M4 MacBook with 16 GB, and your recordings will differ. Treat every transcript as a draft.

Words wrongTwo meetings from the AMI Meeting Corpus, four people each, headset mix. The word error rate was between 21% and 29% for Apple's engine, Parakeet and Whisper large-v3 turbo, and between 29% and 39% for Whisper base, the small optional download. Fillers are left out of the count, and many of the errors are dropped short replies and spelling conventions such as gonna and going.
Speaker labels, meetingsOn the two AMI meetings, counting every word, about 83% to 84% of words got the right speaker, with the standard speaker separation and with Nemotron 3 alike. Some words get no label at all. The speaker setting was tuned on these two meetings, so they may flatter it.
Speaker labels, two-person recordingsFive real two-person recordings from VoxConverse, 71 minutes in all, English, YouTube-quality audio, with labels checked by people. They have no transcripts, so this is scored on speech time, not words. Standard speaker separation missed 8.5% of the speech, added 1.7% that was not speech and mixed up the speakers for 1.1%, a total error of 11.3%. NVIDIA Nemotron 3 missed 1.4%, added 0.5% and mixed up 0.0%, a total of 1.9%. These files were not used to tune the setting. Five files is a small sample, and the app's word accuracy on them was not measured.
Names foundAgainst the corpus's human labels on those meetings, identifier removal found 13 of 15 personal names and 2 of 2 places, and flagged one false place. That is a very small sample.
Scanned textOn 9 generated test files (11 pages), the OCR word error rate was 1.5%.

Privacy and what downloads

Your files stay on your Mac

Transcription, speaker labels, identifier removal, OCR, dictation and draft summaries run on your Mac. There are no accounts. If your ethics board requires that recordings stay on your own computer, Tapefolio does its work there. Whether that meets the board's rules is for you and the board to decide.

Downloads, each once

The internet is used for three things, and none of them carries your files. The optional model downloads inside the app come from Hugging Face, once each. The app's check for a new version sends the app's name, its version, the macOS version and the app's update token to purplelink.llc. And macOS itself may download Apple's own speech model once if you use Apple's speech recognition. There is no analytics and no account. If you connect an AI app such as Claude Desktop, the transcript text it reads goes to that app's servers, with identifiers replaced by codes unless you turn that off. Details are on the privacy page.

Voices are opt-in

Speaker memory is off until you turn it on for a project and record a consent note. What it saves is stored encrypted on your Mac, and you can delete it.

Models
ModelForSize
Apple's speech recognitionTranscription, used until you download a modelNone from Tapefolio. macOS may fetch Apple's model once
NVIDIA ParakeetEnglish transcriptionAbout 470 MB
Whisper baseEnglish transcription, a small downloadAbout 150 MB
Whisper large-v3 turboThe largest model, and the slowest; many languages, and it can detect each interview's languageAbout 650 MB
NVIDIA Nemotron 3 DiarizationSeparating speakers. Recommended, and used automatically once installedAbout 190 MB

Optional models can be deleted from Settings. The first time each model runs on a Mac, it can take a few minutes while macOS prepares the model for the Neural Engine. Whisper large-v3 turbo can take about ten minutes the first time it runs on a given chip.

Credits: the speech models are NVIDIA Parakeet TDT 0.6B v2 (CC BY 4.0), run through FluidAudio and FluidInference (Apache-2.0), and OpenAI Whisper (MIT), run through WhisperKit (MIT, Argmax). The speaker models are from pyannote and WeSpeaker, via FluidInference. The measurements above use the AMI Meeting Corpus (CC BY 4.0) and five recordings from VoxConverse (CC BY 4.0).

Price and trial

Seven days free

The trial is the complete app, counted from the first time you open it. No account and nothing to enter.

$29.99, once

One payment in US dollars, through Stripe. After the trial the app asks for a license key, which arrives in the receipt email and on the page after you pay. Every update is included.

One key, two Macs

The key works in Tapefolio on up to two of your own Macs, works offline and needs no second download. Refunds within 14 days: email ben@purplelink.llc.

System requirements

macOS26 or later.
MacApple silicon. There is no Intel version.
Disk spaceThe app is about 40 MB, plus the optional models you choose to download (see the sizes above).
PermissionsmacOS asks for microphone access the first time you record in the app, and for permission to record system audio the first time you record a call. Dictation types for you only with Accessibility permission.
Draft summaryA Mac with Apple Intelligence turned on. Everything else works without it.

Comparison

How it compares

Tapefolio is one of several Mac apps that transcribe on the device. Here are five of them.

Tapefolio compared with MacWhisper, Aiko, Buzz, Biscotti and Yogurt
TopicTapefolioMacWhisperAikoBuzzBiscottiYogurt
Price$29.99 once, after a 7-day trial.Free tier. Pro is a one-time €64 with lifetime updates.$24 once.Free, MIT license.Free, source available (PolyForm Perimeter license).Free, MIT license.
Where it transcribesOn your Mac. Nothing from your files is sent anywhere.On your Mac with local models. Cloud services are optional.On your device. The page says audio is never uploaded.Offline, on your computer.On your Mac. The page says audio and transcripts are not uploaded.On your computer. The page says audio, transcripts and AI processing stay local; it goes online for model downloads and update checks.
Speaker labelsYes, on the Mac.Listed as a feature.Not supported yet, per its page.Listed as a feature, as speaker identification.Yes, with voiceprints matched across meetings.Names speakers from Zoom, Microsoft Teams and Google Meet, on Mac only.
PlatformmacOS 26 or later, Apple silicon only.Minimum macOS not stated.macOS, iOS and visionOS.macOS, Windows and Linux. Macs need Apple silicon.macOS 15 or later, Apple silicon. 16 GB of memory recommended.macOS 14.4 or later on Apple silicon, Windows 10 and Linux. 8 GB of memory.
Free way to tryThe full app for 7 days.A free tier with no time limit.Not stated.Free.Free.Free, no account.

Competitor facts from macwhisper.com, sindresorhus.com/aiko, github.com/chidiwilliams/buzz (checked on October 13, 2026), github.com/scosman/Biscotti (checked on October 20, 2026), and github.com/FrigadeHQ/yogurt (checked on November 12, 2026). "Not stated" means the page did not say. I did not test the other apps.

Pick MacWhisper or Buzz if you want a free option. Check the MacWhisper requirements first if you have an Intel Mac or an older macOS, because Tapefolio will not run on either. Pick Biscotti if you want free meeting notes with summaries and action items. Pick Yogurt if you want free meeting notes on Windows or Linux as well as the Mac, or on a Mac older than macOS 26. Pick Aiko for the lowest one-time price when you do not need speaker labels. Pick Tapefolio when you want speaker labels, a review step for names and other identifiers, and OCR in one app on your Mac. Prices for more apps, including cloud services, text to speech and OCR, are in the Mac apps for transcription, text to speech and OCR guide.

Frequently asked questions

What does Tapefolio need to run?
macOS 26 or later on a Mac with Apple silicon. There is no Intel version.
Do my recordings leave my Mac?
No. Transcription, speaker labels, identifier removal and OCR all run on the Mac, and nothing from your files is sent anywhere. The internet is used for three things. First, the optional model downloads inside the app come from Hugging Face and happen once each. Second, the app's check for a new version sends the app's name, its version, the macOS version and the app's update token to purplelink.llc. Third, macOS itself may download Apple's own speech model once if you use Apple's speech recognition. There is no analytics and no account. If you choose to connect an AI app such as Claude Desktop, the transcript text it reads goes to that app's servers, with names and other identifiers replaced by codes unless you turn that off.
How accurate is it?
It will mishear words and mislabel speakers. On two meeting recordings from the AMI corpus the word error rate was between 21% and 29% for Apple's speech recognition, Parakeet and Whisper large-v3 turbo, and between 29% and 39% for the small Whisper base model. A person should check any text that will be quoted against the recording. The measured results are above.
Does it anonymize a transcript?
No. Identifier removal finds many identifiers and replaces them with codes, and it misses some. You review each one, and you should read the result before you share it. It does not make a study compliant. That is for you and your ethics board to decide.
What happens to the key file?
When Tapefolio replaces an identifier it writes a separate key file that maps each code back to the original. Anyone who has both the transcript and the key file can reverse the replacement, so store the key file somewhere other than where you share the transcript.
How good are the speaker labels?
Good on clear two-person recordings once Nemotron 3 is installed, and weaker without it. On five real two-person recordings (71 minutes, English, YouTube-quality audio) standard speaker separation had a total error of 11.3% of the speech time and Nemotron 3 had 1.9%. On the two AMI meetings, with four people each, about 83% to 84% of all words got the right speaker, and some got none. Five files is a small sample. Those files have no transcripts, so word accuracy on them was not measured. You can rename and merge speakers in the review screen.
Does it remember people's voices?
Only if you turn speaker memory on for a project, and the app asks for a recorded consent note first. Voices are stored encrypted on your Mac, Tapefolio only suggests matches for you to confirm, and you can delete them. It has been tested only on recordings made the same day with the same microphone, so treat a suggestion as a suggestion.
Which speech model should I use?
Start with the default, Apple's on-device speech recognition, which needs no download from Tapefolio. On the two meetings measured, Apple's engine, Parakeet and Whisper large-v3 turbo were within about two points of each other, so none is proven better. Whisper base made more errors in the same tests. Parakeet was the fastest and is English only. Whisper large-v3 turbo was the slowest. For interviews in another language, use Whisper large-v3 turbo or Apple's engine, since Parakeet and Whisper base are English only.
Why is the first run slow?
macOS prepares a speech model for the Neural Engine the first time that model runs on a Mac. That can take a few minutes for each model, and it happens once for each. Later runs of the same model start at once.
Does one license work on more than one Mac?
One license key works on two of your Macs. Paste the same key into Tapefolio on each one.
Are updates included?
Yes, all of them. The price is a single payment. Tapefolio checks for a new version and offers to install it in place.
What happens when the trial ends?
The trial is the complete app for seven days from the first time you open it, with no account. To keep using Tapefolio after that, buy it. You get a license key by email and on the page after you pay. Choose Enter License Key in Tapefolio and paste it. It works offline and needs no second download.
Can I get a refund?
Email ben@purplelink.llc within 14 days of purchase. See the terms.
How do I get help?
Email ben@purplelink.llc. If you lose your download link, the Find a purchase page emails it again to the address you paid with.
Why not use a free app like MacWhisper or Buzz?
You can, and for a plain transcript a free app may be the better value. MacWhisper has a free tier and Buzz is free and open source, and both transcribe on your Mac. Tapefolio costs $29.99 once. What that pays for is the work around the transcript: speaker labels, a review step for names and other identifiers, and OCR, all on your Mac. Try the 7-day trial on one recording before you decide.
Can it record a Zoom or Microsoft Teams call?
Yes. Record a call puts the call's sound on its own track beside your microphone, and no bot joins the call. Your words are labelled You, and the people on the call are told apart on the call track. You start and stop it yourself. Before each call Tapefolio asks you to confirm that everyone has agreed to be recorded, because recording laws differ and some places require everyone's consent. macOS asks once for permission to record system audio.
Can I ask Claude about my transcripts?
Yes, if you choose to. Tapefolio can act as a read-only MCP server for Claude Desktop and other apps that support MCP: Settings, AI apps shows the line to add. By default the AI app sees names and other identifiers as codes such as [PERSON 1], speakers as Speaker 1 and so on, and never the file names or the key files. The removal can miss some names, and what the AI app reads goes to its own servers, so check a transcript before you share it this way.
Does it work offline?
Yes, once the models you use are installed. Transcription, speaker labels, identifier removal and OCR run on your Mac. The internet is used for the optional model downloads, for the update check, and for the one-time download of Apple's speech model if you use Apple's speech recognition. Pasting a license key works offline.
Is it more accurate than other apps?
Tapefolio has not been compared with other apps for accuracy, so this page makes no claim. The measured results on this page are for Tapefolio's own speech models on two meeting recordings, with word error rates between 21% and 39%. Whichever app you use, check any quoted text against the recording.

Write to Ben

A question about Tapefolio that this page does not answer? Write to Ben here.

Prefer your own mail app? Write to ben@purplelink.llc. The form delivers your message by email through Resend and does not store it on this site. Privacy.

Try it on one recording.

Seven days of the complete app, no account and nothing to enter. Then decide whether to keep it.

macOS 26 or later · Apple silicon