You have a recording of three people around one microphone: a research interview, a hearing, a kitchen-table conversation you will quote from. Drop it into any transcription tool and the words come back in about the time it takes to make coffee. Transcription, in 2026, is solved.

What is not solved is the line above each paragraph. Without it you get a wall of text in which a question and its answer are indistinguishable, so you re-listen to the audio anyway — the thing you were trying to avoid. And there is a second problem almost nobody names: next week you record the same person and they are Speaker 2 all over again, because the tool threw away what it learned about their voice the moment the file finished.

This guide covers what the job really consists of, the four honest ways to do it, and what each costs you — in money, in setup, and in who ends up holding a recording of your interviewee’s voice.

Disclosure: I make Samurai Scribe, one of the tools below. Every claim about the others comes from their own pricing pages, documentation or repositories, checked on 21 October 2026 and linked where I make it. Prices and features in this category change monthly; check before you buy anything.

Three problems wearing one name

“Speaker labels” is shorthand for three separate pieces of machinery. Tools advertise whichever ones they have in the same words, which is why comparing them is harder than it should be.

Transcription. Audio to words. Whisper, Parakeet and their descendants do this well enough that it is rarely what fails.
Diarization. Cutting the audio where the voice changes and grouping the pieces, so this recording has a Speaker 1 and a Speaker 2. This is the part that struggles, and where most of the errors you will curse at come from.
Identification. Deciding that Speaker 2 is Alice, the same Alice from the last four recordings. This needs a stored voice print — a numerical fingerprint of how someone sounds — and somewhere to keep it.

Most tools do (1). Many do (2). Few do (3), and those that do it well mostly do it in their cloud, which is a choice with consequences for a recording of someone who agreed to talk to you. Here is one recording after each stage:

Most tools stop after stage two. Stage three is the one that decides whether next week's recording arrives already named.

What to check before you trust a tool with an interview

Five questions, in the order they will bite you:

Does it handle one microphone with several people? Many “speaker label” features really separate audio sources — your mic versus the call’s output. That gives you “me” and “them”, never three people at a table.
Where does the audio go? If the answer is a server, your interviewee’s recording is in a second organisation’s hands, whatever the retention policy says. There is a one-minute test that settles this for any tool, including mine: turn the network off and record. Either a transcript appears or it does not.
Does it remember a voice between recordings? Irrelevant for a one-off; for a longitudinal study it is the difference between ten seconds of tidying per session and ten minutes.
If it does, where is the print and can you delete it? A voice print is biometric data about someone who is not you.
Can you fix a wrong label everywhere at once? You will need to. Plan for the machine being wrong.

Route one: a human transcription service

Worth saying first, because for some of this work it is simply the right answer. Rev charges $1.99 per minute for human transcription, claims 99%+ accuracy and returns work in 12 hours or less; legal transcripts are sold per page, at $2.50 for a Legal Premium Transcript with a claimed 99.7%+ accuracy.

If the transcript will be quoted in a published paper, filed, or relied on by someone other than you, a human is still better than every machine in this article, mine included. The objections are cost, the wait, and a stranger hearing your recording — an hour of audio runs about $120, which is fine for one important interview and impossible for thirty.

Route two: a cloud tool

The easiest route, and the one where identification is most mature. Otter.ai genuinely does the hard part: its own help centre says that tagging a speaker “will teach Otter to recognize that speaker and improve upon the speaker identification for your future conversations”, and in a Workspace those profiles are shared between members, so a colleague you have never tagged can be named automatically in your recording. That is a real feature and I will not pretend otherwise.

But read the rest of the same page. The FAQ answers “Can I delete a speaker?” with: “No. You cannot delete a speaker at this time.” So the voice profile of the person you interviewed lives in Otter’s account system, may be shared across your workspace, and cannot be removed. The free tier is 300 minutes a month, capped at 30 minutes per conversation; Pro is $16.99/month, or $8.33 billed annually, for 1,200 minutes and a 90-minute cap. Those caps bite, because interviews are long.

A federal judge also allowed wiretap and Illinois biometric-privacy claims about voiceprints against Otter to proceed in August 2026; nothing has been proven, and I cover the ruling in the Otter.ai alternative guide.

Descript is the other common answer, and it detects 8+ speakers on every tier, including the free 60 minutes a month. Its security page says uploaded files sit on Descript’s servers and in cloud storage, with Whisper “hosted in Descript’s infrastructure” — and that sharing transcripts to improve the service is opt-in and off by default, which is better than the norm here.

Route three: build it yourself

If you are comfortable at a command line, WhisperX is free, BSD-2-licensed, and does transcription, word-level alignment and diarization in one pipeline. Diarization comes from pyannote-audio: you need a Hugging Face access token and you must accept the user agreement for the speaker-diarization-community-1 model. It runs on CUDA, or on CPU with --compute_type int8 --device cpu, which is also the suggested path on macOS.

This is the best option if you are on Linux, want it free and auditable, or are processing a hundred files in a script. Two caveats. The README calls its own diarization “far from perfect” — the same models everyone else uses, with the same failure modes. And it solves problem (2), not problem (3): recognising the same person across recordings is a separate job this pipeline does not do, and pyannoteAI sells that as a hosted API instead.

Route four: a local app

Everything runs on your machine: no token, no CLI, no upload. Checked 21 October 2026 from each product’s own pages.

	Price	Platforms	Audio goes to	Several people, one mic	Same person next time
Samurai Scribe	Free · Pro $19 once (beta; $39 list)	Windows, Apple Silicon Mac	Your computer	Yes, free tier	Yes, prints in your folder
MacWhisper	Free · Pro €64 once	Mac	Your Mac; speakers need WhisperKit or a cloud API	Yes	Autofill, not voice matching
Meetily	Free (open source) · Pro $120/yr	Windows, macOS; Linux waitlist	Your computer	Diarization is Pro	Not stated
WhisperX	Free, BSD-2	Windows, macOS, Linux	Your computer	Yes	No
Otter.ai	Free 300 min/mo · from $8.33/user/mo	Web, Mac, Windows, mobile	Otter’s cloud	Yes	Yes, in Otter’s cloud

One row needs explaining, because it is the row I would have got wrong while shopping. MacWhisper is a very good app and its automatic speaker recognition is real, but its documentation describes labels within one transcript — “Speaker 1, Speaker 2 etc” — renamed and merged by hand. The cross-file part is autofill: “autosuggestions based on the speakers you’ve created in settings will appear.” That is help with typing, not matching by voice. It also needs the WhisperKit engine or the cloud providers ElevenLabs and Deepgram.

If you are on a Mac and transcribing files you already have, MacWhisper is excellent and I would not try to talk you out of it. If open source is your deciding factor, Meetily and WhisperX are the honest answers; mine is not. What I built differently is Windows as well as Mac, diarization free, and voice prints kept as files in your own folder.

Doing it, step by step

1. Record the room, or import what you have

Install it — Windows 10/11 64-bit or an Apple Silicon Mac, no account, no email, about 1.5 GB for the app and its models. Audio is captured at the operating-system layer, so a call is recorded with nothing joining it and a conversation at a table works the same way. Interviews already on a phone import identically.

2. Let it find the voices

One microphone carrying three people is split where the voice changes, grouped into speakers, and each speaker gets a voice print. The models are open ones you can look up: pyannote’s speaker-diarization-community-1 for segmentation — the same model WhisperX asks you to download — and 3D-Speaker ERes2Net for the embeddings. No secret sauce, and anyone claiming some should be asked for numbers. Mine are further down.

3. Name each person once

A Speakers rail beside the transcript lists who spoke and for how long. Play a few seconds, type a name, drag one speaker onto another to merge them when the machine split one person in two. If the recording came from a calendar invite, the invitees are offered.

4. Watch the second interview name itself

This is the part that changes the work. Interview one costs you twenty seconds of naming. Interview five, with the same two people, can open already named — and over a study of thirty sessions that is the difference between a chore and nothing at all.

Once you confirm a name, that person’s voice print is kept — in your vault, as a file — and later recordings are matched against it. A strong match is labelled automatically and flagged unconfirmed, so you can see what was guessed; a weaker one is offered (“sounds like Bob”) rather than asserted; below that the speaker stays Speaker 2 and nothing pretends otherwise. That ladder exists because, as the numbers below say, the machine is not always right, and a confident wrong label costs more than an honest blank one. Since v0.62.0 it also happens live, as you record.

You can delete a voice print, or clear it and keep the person. It is a file in a folder you chose.

The print is a file in a folder you chose, not a profile in an account. That is the whole difference between this and doing it in the cloud.

5. Keep the result as files

Every recording becomes a folder of Markdown: a note, a transcript, a summary, the audio. Transcript segments are headings — ### 12:13.4-12:15.0 [[Alice Chen]] — and each named person gets People/Alice Chen.md, so in Obsidian every person becomes a page with backlinks to every conversation they appear in. For a researcher coding a dozen interviews that is an index for free; more on the layout in meeting notes in Obsidian.

Where this is still wrong

It would be easy to stop at the last section, so here is the part that does not fit in a feature table.

Diarization is the weak link, in every tool including mine. Across seven of our own test recordings the diarization error rate ran from 13% to 48%, median near 22%. Four of the seven sat between 13% and 22%; two were in the forties, and both of those were ordinary two-person recordings, not exotic ones. I am giving you the spread rather than the best number because the best number would mislead you.

Four of the seven sit between 13% and 22%. The two in the forties are ordinary two-person recordings, not exotic ones — which is why the spread is here instead of the best number.

What it got right in all seven was how many people were in the room; what it gets wrong is where one stops and the next starts. So you rarely have to invent a missing speaker — you fix boundaries and the odd swapped line. Which is why the thing to judge is not whether a tool is ever wrong but what being wrong costs you: renaming a speaker rewrites every segment, and merging two the machine split is one drag. That is a deliberate answer to a problem nobody has solved.

Identification tested better — with a caveat I have to state. Matching a voice across different recordings came out near 0.4% equal error rate, against 129 pairs of different people. But the same-person side of that test was six pairs, all of one voice: mine. Enough to pick this embedding model over five rivals, which is what the test was for. Not enough for me to promise you a number, so I won’t.

Two users have told me in-room speaker identification is imperfect for them, on a Teams call and at a table. Two reports is not a trend, but it is the clearest product gap I have, and I would rather you heard it from me than found it on your own recording.

Recording other people

Removing the cloud removes a third party from your recording. It does not remove your own obligations. Consent rules differ by state and country — some require everyone’s agreement, some only one person’s — and a research interview carries a separate duty to the person in front of you whatever the law says. Tell people they are being recorded. A voice print is data about them, so being able to delete one when you are asked is worth something. This is not legal advice.

FAQ

Does it work with no internet? Yes, once the first run has downloaded a model. Transcription, diarization, voice prints and summaries all run on your CPU or GPU. The app still makes a daily update check and, with Pro, a licence check; both are on the network disclosure.

How many people can it tell apart in one room? No fixed cap, and accuracy falls as voices get closer together and people talk over each other. Three or four distinct voices around one microphone is the case it is built for. Phone and handheld recordings import and get the same treatment.

Where is the voice print kept, and what is in it? In your vault, as a small file per person: a numerical summary of how someone sounds plus which confirmations taught it, not audio. You can clear it and keep the person, or delete the person entirely.

Is any of this behind the paywall? No. Speaker detection, voice prints and the People notes are all free, on a tier with no time limit and no minute cap, running Whisper Tiny, Base or Small, or NVIDIA Parakeet. Pro — $19 once during the beta, $39 after — adds the larger Whisper models and Parakeet Ultra: more accurate words, not better speakers. To hear the difference on your own interview audio, download a Pro model in Settings → AI Models → Transcription and try it in Compare models.

Can I use this for confidential interviews? I can’t answer that for you, and anyone who does is overreaching — whether a recording may be made and where it may be processed is a rule about you and whoever you answer to. What I can tell you is what the software does: audio, transcription, speaker work and summaries (on the default local model) stay on your machine. The connections it makes on its own are few: a daily update check you can switch off, its model downloads and, with Pro, a licence check, all itemised on the network disclosure. Verify that rather than believe it — pull the network and record.

What about Linux? Not yet. WhisperX is today’s answer there; Meetily has a waitlist.

So should you try it?

Not everyone should, and the honest version is short.

Use a human service if the transcript will be quoted, filed or relied on by someone else. Rev at $1.99 a minute buys accuracy no local model here will match.

Roll your own with WhisperX if you are on Linux, want it open source, or are processing files in bulk. You will not get cross-recording identification; everything else is free.

Stay with a cloud tool if you need a shared archive your colleagues can search, and your recordings are the kind you are comfortable keeping on someone else’s servers. Plenty are. Local-first tools are genuinely worse at team archives.

Try this if your recordings happen in a room rather than on a call; if you record the same handful of people repeatedly and are tired of naming them; if you are on Windows and keep finding the good local tools are Mac-only; or if the recording is of someone who trusted you with it.

The test takes ten minutes, and you should run it with the wifi off — that way you are checking both claims at once. Record five minutes with two colleagues, name them, record again, and see whether the second one names itself. That answers the only question that matters here better than this article can, and what you are left with either way is a folder of Markdown and audio that is yours.

An Otter.ai alternative that works offline — what the August ruling actually said, and what moving off a cloud notetaker involves.
Meeting notes in Obsidian, transcribed on your own computer — the vault layout, person notes and what agents can do with them.
A MacWhisper alternative for Windows — if you are on a PC and keep hitting Mac-only tools.
How to transcribe YouTube videos and lectures offline — for files and links rather than live conversations.