Transcription with speaker labels
7 services split a recording into speakers, and every row below says where we read it. Transcrib Si separates the voices on every file of every plan, the free one included, and the name stays on the line when you export.
Why the split matters
A recording of two people talking turns into a wall of text the moment nobody writes down who said what. The reader has to guess whether the second question belongs to the same person as the first, and in a job interview or a long meeting that guess is the whole point of the recording. So every service on this list works on the same small problem: hear the change of voice, cut the text there, and put a name on the piece.
The word for that work differs from house to house, and the promises differ too. One service calls it diarization, another calls it speaker recognition, a third writes about an algorithm that tries to identify the speakers and then admits, in the same paragraph, that it is not perfect. We read 7 services on 25 October 2026 and wrote down what each of them actually claims, with the address of the page and the date next to every row. Where a page says nothing about the feature, the table says that too, instead of guessing.
| Service | What it says about the speakers | Is it in the free part? | Where we read it |
|---|---|---|---|
| Transcrib Si | one line per segment, with the time it starts and ends and the speaker it belongs to; the export files carry the same structure | yes: the free tier is $0 with 3 files a day of up to 30 minutes, and speaker labels are on every file of every plan (pricing, checked ) | the general guide, checked |
| Sonix | it detects when the speakers change and separates the dialogue into labeled paragraphs, and you can put real names on them from a dropdown | the price page says every plan includes the editor, and a new account gets 30 minutes of free transcription (sonix.ai/pricing, checked ) | sonix.ai, automated transcription, checked |
| TurboScribe | a Speaker Recognition checkbox under More Settings when you upload; the page says it takes an extra minute or two per hour of audio to get a transcript labeled with speakers | the source does not say which plans carry the option, and it names 3 free transcripts a day (turboscribe.ai, pricing, checked ) | turboscribe.ai, pricing, checked |
| HappyScribe | automatic speaker detection on the price list, and speaker diarization on the home page, where it says the split tells people apart on real meetings | yes: the Free plan lists automatic speaker detection, with 45 minutes per recording (happyscribe.com/pricing, checked ) | happyscribe.com, home page, checked |
| Fireflies | speaker recognition, described as identifying different speakers in meetings and audio files | the page we read does not say whether the free plan carries it (fireflies.ai, home page, checked ) | fireflies.ai, home page, checked |
| Trint | automatic speaker recognition and time codes, listed among the metadata the workflow gathers as the recording goes on | the source does not report a free tier on the pages we opened (trint.com, home page, checked ) | trint.com, home page, checked |
| Temi | an algorithm tries to identify the speakers and separate them into paragraphs, and Temi says in the same breath that it is not perfect; you can rename a speaker and choose whether the export carries the names | the first file is free up to 45 minutes, and no separate charge is named (help.temi.com, Accounts and Billing, checked ) | help.temi.com, Editing Transcripts, checked |
The words each service uses
The feature is sold under at least four names, and the names say something about the promise behind them. A label is something a person types. Recognition is something the machine hears. Diarization is the technical word for cutting a stream of speech into turns. Detection is the word for finding the change at all. The table below is only about the wording, so that a search of the page you are on finds the same word the service itself uses.
| Service | The term on their own page | Where we read it |
|---|---|---|
| Transcrib Si | speaker labels | pricing, checked |
| Sonix | automated speaker diarization | sonix.ai, automated transcription, checked |
| TurboScribe | speaker recognition | turboscribe.ai, pricing, checked |
| HappyScribe | automatic speaker detection, and speaker diarization on the home page | happyscribe.com/pricing, checked |
| Fireflies | speaker recognition | fireflies.ai, home page, checked |
| Trint | automatic speaker recognition | trint.com, home page, checked |
| Temi | a speaker identification algorithm | help.temi.com, Editing Transcripts, checked |
How ours works, step by step
- One switch. The uploader has a Split by speaker switch next to the language selector. The API takes the same switch under the name diarize, set to true.
- The result is a transcript, not a guess. Every segment carries the time it starts, the time it ends and the speaker it belongs to, so the reader can jump to the moment and hear the voice again.
- The name survives the export. The export files carry the same structure as the transcript: one line per segment, with its time and its speaker. TXT, Markdown, SRT, VTT and CSV on every plan, DOCX on Pro.
- The free tier has it too. Speaker labels are not a paid extra here. The free plan is $0, 3 files a day, up to 30 minutes per file.
- Two languages. The split runs over Russian and English, and the language is detected when you do not set it. We do not claim more than the two, because the contract says two.
What this page does not tell you
No service on this list has published a test that puts all of them on the same recording, with the same speakers, and counts the wrong labels. So this page cannot tell you which split is the most accurate, and we are not going to pretend otherwise with a ranking. What it can do is tell you what each service claims in its own words, where that claim stands, and whether the claim survives contact with the free part of the plan.
Two rows are worth reading twice. Fireflies says on its home page that it identifies different speakers in meetings and audio files, and the page we read says nothing about the free plan. Trint collects automatic speaker recognition and time codes as metadata inside a wider workflow, and the pages we opened named no free tier at all. Temi is the honest one: it says the algorithm is not perfect, which is the most useful thing a service can tell you before you hand it two hours of a difficult conversation.
One more caution about the numbers you will meet elsewhere. A percentage of accuracy and a percentage of correctly separated speakers are two different measurements, and pages often print the first next to the second. We checked every row against the sentence that carries it, and where a service publishes a percentage we did not find, the row says the source does not report one.
Turn the switch on and read the names.
The widget below is the same one as on the home page: file → transcript with timecodes and speakers → export. The transcript is saved to your account.
Drop an audio or video file
or — mp3, wav, m4a, mp4, up to 64 MB and two hours
Transcription runs under your account: sign in or create a free account.
No account yet? The free plan needs no card. Create a free account or sign in.
Questions this page answers.
Which service separates the speakers best?
No service has published a test that puts all of them on the same recording and counts the wrong labels, so we cannot rank them and we do not try. What we can do is show what each service claims in its own words, with the page and the date, and whether the claim is on the free plan.
Does Transcrib Si label the speakers on the free plan?
Yes. Speaker labels are on every file of every plan, and the free tier at $0 gives 3 files a day of up to 30 minutes each with the labels included. The split is a switch in the uploader called Split by speaker, and the API takes the same switch as diarize set to true.
Do the export files keep the speaker names?
They do. The export carries the same structure as the transcript: one line per segment, with the time it starts and ends and the speaker it belongs to. TXT, Markdown, SRT, VTT and CSV are available on every plan, and DOCX comes with Pro.
What is the difference between a label and a recognition?
The wording says what the service promises. A label is something a person types onto a piece of text. Recognition is something the machine hears on its own. Diarization is the technical name for cutting one stream of speech into turns, and detection is the name for finding the change at all. Temi writes that its algorithm is not perfect, which is the most useful kind of promise to read before a long conversation.
Which languages do you recognise?
Russian and English, and the language is detected on its own when you do not set it in the uploader. We do not claim more than the two, because our own contract says two. Every row in the table above names the language count of the other service and the page it was read from.
Related pages.
The percentages each service publishes, and where.
A workspace with no published price.
A bot that joins the meeting.
Paying by the audio minute.
A panel inside the browser.
A suite that sells minutes by the pack.
A checklist for calls with speakers and timecodes.
The general guide to the whole service.
Two hours of audio and how to get it through.
Mark Burmistrov, Founder.
Every number on this page comes from a public page we opened ourselves, from the service contract or from our own measurements. The author of the page and the publisher of the site are the same person. Questions and corrections: support@transcrib.si.