Your audio never leaves your device
Free transcription with speaker labels, so you can see who said what
Speaker labeling transcription that separates voices automatically and shows each person's turns. Rename the speakers to real names, and export the result. Free transcription with speaker labels, in your browser, with no subscription and nothing uploaded.
Takes a little extra time, and your audio still never leaves your device. Labels go by how voices sound, so treat them as a best effort: one person can be split across two or three labels, similar voices can be merged, and someone joining partway through may be folded into a speaker already talking. Giving a number usually helps. Check them before relying on them.
Specifying this number improves accuracy. Count people who really take part, whenever they join. Someone who only says a word or two is better left out.
Timestamps each word instead of each phrase, so short interjections like "Yeah" go to the right person. On a busy four-person recording this corrected about 1 word in 5; it changes little when people talk in long turns. Adds roughly 40% to the time on a computer, and almost nothing on a phone.
Phone audio loses the high frequencies that separate consonants, so they are put back before transcribing. Only phone-quality recordings are treated, and you are told when it happens. Also even out volume is for one case: someone far from the microphone and much quieter throughout. It cost accuracy on other recordings, so use it only if a quiet person came out badly.
Worth filling in. The speech model knows ordinary English, not your subject, so a word it has never heard becomes whatever sounds closest, and it usually gets that word wrong every single time it comes up. A sentence is enough, and the names and jargon in it are picked out and watched for. It cannot guess a word you do not mention, so name the people and the terms that matter.
Free, no signup. The first time you use it, it takes a minute or two to set up. After that setup is quick. Works with MP3, WAV, M4A, OGG, FLAC and AMR, and long recordings are handled in parts.
Who said what, without a cloud service
Speaker labels turn a wall of text into a readable conversation. Without them, a two person interview reads as one long monologue and you have to reconstruct the turns from context. With them, you can scan for one person's contributions, find the moment somebody committed to something, and quote accurately.
Many transcription services charge extra for this or require an upload to get it. Here it is free and runs on your device, using a second model that loads alongside the transcriber.
Diarization is not voice identification
These get conflated constantly, and the difference matters for anyone handling sensitive audio.
- Diarization, which is what this does, answers "how many different voices are in this recording, and which parts belong to each?" It works entirely within one file. Nothing it produces carries over to another recording, and the output is SPEAKER_00 and SPEAKER_01 until you rename them.
- Speaker identification, which this deliberately does not do, answers "whose voice is this?" by comparing against stored voiceprints. That requires a biometric database of known people, and it carries a completely different set of privacy and legal implications.
Being precise about the mechanism matters here, so: in Accurate mode the tool does compute a numeric voice signature for short excerpts of your recording. That is how it recognizes that the person speaking at minute two is the same person speaking at minute forty. Those signatures exist only in your browser's memory while the transcript is being produced, are used only to compare excerpts of that one recording against each other, and are discarded when it finishes.
What that means in practice: nothing is stored on disk, nothing is transmitted anywhere, no signature is ever compared against a different recording, and there is no database of known voices for one to be checked against. The link between SPEAKER_00 and a real person exists only in the transcript you type it into. If you would rather no voice signature be computed at all, Fast mode skips that step entirely, at the cost of accuracy described below.
Choosing a mode
There are three settings, and the right one depends mostly on how long the recording is:
- Accurate. The default, and the one to use for anything longer than a couple of minutes. It compares voices so that one person keeps one label from beginning to end. It downloads a small extra model once, about 6 MB, and makes a second pass over the audio.
- Fast. Segmentation only. It detects that the speaker changed but cannot tell whether a voice returning later is somebody already labelled, so long recordings tend to split one person across several labels. Reasonable for a short clip.
- Off. No speaker processing at all. Fastest, and the right choice for dictation, narration, or any single-voice recording, where labels add nothing.
Set the speaker count if you know it.Leaving it on auto makes the tool decide how many distinct voices exist, which is the hardest part of the job. Telling it "two speakers" removes that guess entirely and is the single most effective thing you can do for label quality.
From labels to real names
- Choose a speaker mode and transcribe as usual.
- The transcript shows generic labels like SPEAKER_00 and SPEAKER_01 on each turn.
- Type real names next to those labels and apply them, and the whole transcript updates at once.
- Export the labeled transcript as text, or as SRT and VTT with names on each caption.
One thing to know about the order of operations: applying names rebuilds the transcript from the original result, which replaces manual text edits you made beforehand. Rename the speakers first, then do your wording corrections, and nothing gets lost.
Why separation goes wrong, and what to do
Speaker separation fails in recognizable patterns. Knowing which one you are looking at tells you whether to re-record, adjust settings, or just fix it by hand:
- Two people merged into one label. Usually similar voices, similar pitch, or one microphone far from both. Hand correction is the only fix within a finished recording.
- One person split across several labels. Typically caused by a change in volume or position partway through, for instance someone leaning toward the microphone or moving across the room. Rename both labels to the same person and the turns merge.
- Labels swap in the middle. The classic symptom of a stretch of overlapping speech. Look for the point where they argue or interrupt, and correct from there.
- A speaker disappears entirely. Usually a quiet participant on a distant microphone. Their words are still in the transcript, attached to whoever was loudest nearby.
- Separation did not run at all. You get a notice and a complete unlabeled transcript. Re-running with a shorter file often succeeds where a long one did not.
Each of these has a different remedy, and one of them is now largely solved: short interjections landing on the wrong person are a timing problem rather than a voice problem. See how to correct wrong speaker labels for what fixes each case, with measurements from a real four-person recording. For how the labels are written into transcripts, SRT and VTT, see the speaker label format guide.
Recording so that separation works
Almost all diarization quality is decided at recording time, not at transcription time. The highest-value habits:
- Equalize distance. Put the microphone between people rather than beside one of them. Uneven levels cause more label errors than any other single factor.
- Separate microphones beat a shared one, if you have the option. Remote calls recorded per participant separate almost perfectly.
- Encourage clean turns. A conversation where people wait half a second before replying separates dramatically better than one with constant overlap.
- Kill steady background noise. Fans, air conditioning, and traffic blur the boundaries the model looks for.
- Have people say their name at the start. It makes matching labels to people trivial when you come back to the file later.
When to turn it off
Separation is not free: it is a second model download and a second pass over the audio, so it makes every transcription slower. Turn it off when the recording is a single voice, such as dictation, narration, a voice memo, or a solo presentation. You get a faster result and a cleaner transcript with no spurious labels to strip out. It is also worth turning off for crowded recordings where you already know the labels will be wrong and you intend to attribute turns yourself.
The mechanics of both models, and how the text is matched back to speakers, are described in how it works.
Last reviewed: August 2026
Frequently asked questions
- How does speaker separation work here?
- Two models run in your browser alongside the transcription. The first slides over the audio and marks when someone is speaking, including when two people overlap. In Accurate mode a second model then compares short excerpts of each voice so the same person keeps one label across the whole recording. Each piece of transcribed text is matched to whichever speaker stretch overlaps it most.
- What is the difference between Accurate and Fast?
- Fast runs only the segmentation model. It knows somebody changed, but it cannot tell whether the person speaking at minute forty is the same one from minute two, so a long recording tends to split one person into several labels. Accurate adds voice matching, which fixes exactly that, at the cost of a one-time 6 MB model download and a second pass over the audio. For anything longer than a few minutes, Accurate is worth it.
- Does it recognize who people are?
- No. It tells voices apart within one recording and nothing more. Accurate mode computes a numeric voice signature in memory to compare excerpts of that recording against each other, then discards it. Nothing is stored, nothing is transmitted, no signature is compared against another recording, and there is no database of known voices. The labels are anonymous until you type real names in yourself.
- How many speakers can it handle?
- Accurate mode handles small groups well and degrades gracefully beyond that. Heavy crosstalk and noisy audio make separation harder, which is true of every automatic system. If you know how many people are in the recording, setting the speaker count instead of leaving it on auto is the single biggest accuracy improvement available.
- Is it perfect?
- No automatic speaker labeling is. Similar voices, overlapping speech, and uneven microphone distance all cause mislabeling. You can correct both the labels and the text in the in-page editor before exporting.
- What happens if separation fails?
- You still get a complete transcript, just without speaker labels, and a notice explaining that separation did not run for that file. The transcription and the separation are independent steps, so one failing does not lose the other.