Transcribe Audio to Text — Free, With Timestamps and Speaker Labels

Upload any audio or video file and get a transcript with speaker labels and timestamps in minutes — 17 formats, 99 languages detected automatically, and a copy you can download before you decide anything.

Drop your audio file here

or click to browse

Audio: MP3 · WAV · M4A · FLAC · OGG · AAC · AIFF · WMA · AMR · OPUS

Video: MP4 · MOV · AVI · MKV · WebM · FLV · WMV

Up to 200 MB · first 5 minutes free, no account

Free 5-minute preview · no account · files up to 200 MB · sign up free for 30 minutes and 5 GB files

What you get back

Every line carries a timestamp and a speaker label. Nothing is collapsed into a wall of text, and nothing is stamped once at the top and left at that. The examples below are written to show the shape of the output rather than taken from any one recording — upload a file above and you will get this structure back.

Transcript — exampleEN · 3 speakers · 214 lines
  1. 00:04Speaker 1So the first thing we need to cover is the quarterly numbers.
  2. 00:09Speaker 2Right — and that ties directly into the hiring question.
  3. 00:14Speaker 1Exactly. Let me pull those up.
  4. 00:18Speaker 3Before you do — are we using the revised forecast or the original?
  5. 00:23Speaker 1Revised. The original was built before the Q2 numbers landed.
  6. 00:29Speaker 3Good. The original had us about eleven percent under on services.
  7. 00:36Speaker 2That is the gap I keep running into when I plan headcount for next quarter.
  8. 00:43Speaker 1Then let us settle the forecast first and come back to headcount after.

Showing the first 8 of 214 lines

Speakers are separated automatically — you can rename them in the editor and the change applies across the whole transcript. Word-level timings for every individual word are in the JSON export, covered below.

The same example, translated

Translation runs on the transcript you already have — no re-uploading, no re-processing. Timestamps and speaker labels stay attached to the right lines, which is the part most translation workflows lose. Here it is in Spanish; 99 languages are available.

  1. 00:04Speaker 1So the first thing we need to cover is the quarterly numbers.
    00:04Speaker 1Lo primero que tenemos que tratar son las cifras trimestrales.
  2. 00:09Speaker 2Right — and that ties directly into the hiring question.
    00:09Speaker 2Cierto, y eso está directamente relacionado con el tema de la contratación.
  3. 00:14Speaker 1Exactly. Let me pull those up.
    00:14Speaker 1Exacto. Déjame buscarlas.
  4. 00:18Speaker 3Before you do — are we using the revised forecast or the original?
    00:18Speaker 3Antes de que lo hagas, ¿estamos usando la previsión revisada o la original?
  5. 00:23Speaker 1Revised. The original was built before the Q2 numbers landed.
    00:23Speaker 1La revisada. La original se elaboró antes de que llegaran los datos del Q2.
  6. 00:29Speaker 3Good. The original had us about eleven percent under on services.
    00:29Speaker 3Bien. Con la original íbamos un once por ciento por debajo en servicios.
  7. 00:36Speaker 2That is the gap I keep running into when I plan headcount for next quarter.
    00:36Speaker 2Esa es la brecha con la que me encuentro una y otra vez al planificar la plantilla del próximo trimestre.
  8. 00:43Speaker 1Then let us settle the forecast first and come back to headcount after.
    00:43Speaker 1Entonces fijemos primero la previsión y volvemos luego a la plantilla.

Showing the first 8 of 214 lines · 98 more languages available

And what a summary looks like

With a free account, the same transcript can be summarized in one click. The format follows the recording — meeting, sales call, lecture, or podcast.

AI summary — exampleFree account

Overview

A 22-minute planning call between three participants covering the quarterly figures and how they constrain next-quarter hiring. The group works through a discrepancy between two forecasts and agrees an order of business before closing.

Key points

  • Quarterly numbers are the first agenda item, and they feed directly into the hiring decision.
  • The revised forecast supersedes the original, which predated the Q2 results.
  • The original forecast ran about eleven percent under on services, which is the gap blocking headcount planning.
  • The group agreed to settle the forecast before returning to headcount.

Action items

  • Pull up the quarterly figures for review00:14
  • Confirm everyone is working from the revised forecast00:23
  • Resolve the services shortfall before planning next-quarter headcount00:43

How to transcribe audio to text

Three steps, and the first five minutes need no account.

  1. 1

    Upload the file

    Drag an audio or video file into the box at the top of this page, or click to browse. Seventeen formats are accepted and the audio track is pulled out of video automatically — no converting first.

  2. 2

    Wait for the transcript

    The language is detected from the audio, speakers are separated, and every line is timestamped. A one-hour recording takes about 5-10 minutes; the free preview covers the first five minutes and comes back in well under one.

  3. 3

    Edit, then export

    Fix any misheard names in the editor and rename speakers — a rename applies to every line at once. Then download: TXT and SRT without an account, or DOCX, VTT, CSV and JSON once you have one.

Timestamps: segment, word-level, and subtitle cues

“Timestamped” means three different things depending on what you are doing with the transcript, and the right one differs by job.

Segment-level

A stamp at each speaker turn, every few seconds. This is what you read on screen and what the DOCX and TXT exports carry. Right for reading along, quoting a passage, or dropping chapter markers into show notes.

Word-level — in the JSON export

Every individual word timed to the millisecond, with a confidence score. Needed for developer pipelines, importing into NVivo, ATLAS.ti or MAXQDA, and verifying a quote against the exact moment it was said. Available in the JSON export with an account.

Subtitle cues — SRT and VTT

Subtitle files are not one cue per word; that would be unreadable on screen. Cues cap at 80 characters and 6 seconds, and always break on a speaker change or a silence longer than a second — so lines fit the frame and stay up long enough to read.

What the word-level data actually looks like

Every word carries its own start time, end time, and a confidence score between 0 and 1. Nothing else on this page's search results shows the structure, so here it is — this is the shape the JSON export returns, one object per word.

{
  "segments": [
    {
      "start": 4.12,
      "end": 7.94,
      "speaker": "Speaker 1",
      "text": "So the first thing we need to cover is the quarterly numbers.",
      "words": [
        { "word": "So",        "start": 4.12, "end": 4.24, "probability": 0.99 },
        { "word": "the",       "start": 4.24, "end": 4.33, "probability": 0.99 },
        { "word": "first",     "start": 4.33, "end": 4.58, "probability": 0.98 },
        { "word": "thing",     "start": 4.58, "end": 4.81, "probability": 0.99 },
        { "word": "quarterly", "start": 6.90, "end": 7.42, "probability": 0.94 }
      ]
    }
  ]
}

The confidence score is the useful part in a pipeline: it lets you flag the words worth a human glance instead of re-reading everything. Proper nouns and technical terms are where it drops, which is exactly where transcripts need checking anyway.

For timestamp formats, standards, and tooling in general, see how transcription timestamps work.

Everything the transcript workspace does

The transcript is the starting point, not the deliverable. Once a file is processed, everything below runs on it without re-uploading or re-processing.

  • Timestamps on every transcript and every export
  • Automatic speaker labels — multi-speaker, renameable in the editor
  • Word-level timestamps in the JSON export
  • Translate the transcript in place — free translations included; premium keeps terminology consistent across long recordings
  • AI chat — ask questions, answers cite the timestamp
  • AI summary — meeting, sales, lecture, and podcast formats
  • Exports: TXT, SRT, VTT, DOCX, CSV, JSON
  • Bulk upload and bulk export
  • Editor with audio-synced playback
  • 99 languages, detected automatically

Free previews run on our premium speech recognition model. Signing up adds minutes, larger files, and the AI workspace — translation, chat, and summaries.

Formats and file sizes

Ten audio and seven video formats. For video, the audio track is extracted automatically — you do not need to convert anything first.

Audio

  • MP3Most common
  • WAVLossless
  • M4AiPhone default
  • FLACLossless
  • OGGOpen format
  • AACApple/streaming
  • AIFFPro audio
  • WMAWindows
  • AMRMobile
  • OPUSModern web

Video

  • MP4Most common
  • MOVApple/QuickTime
  • AVIWindows legacy
  • MKVHigh-quality
  • WebMWeb video
  • FLVFlash legacy
  • WMVWindows

Free preview: files up to 200 MB, first 5 minutes transcribed, no account. Free account: 30 minutes, files up to 5 GB and 10 hours long.

For comparison: Canva's free converter caps at 4.5 MB — about four minutes of MP3 — Zamzar at 200 MB, and the Whisper API at 25 MB. Multi-hour lectures and uncompressed WAV masters have to be split before most converter sites will take them.

How accurate is it — measured, not claimed

95%+ on clear English audio with a single speaker — roughly a 5% word error rate. That number travels with its condition, because accuracy on real recordings varies more than any single figure suggests.

  • Clear podcast audio: 3–6% WER (94–97% accurate)
  • Noisy interviews, background music: 8–15% WER (85–92% accurate)
  • Strong accents, technical jargon, overlapping speakers: 10–20% WER (80–90% accurate)

Review transcripts before publishing critical content. No AI tool reaches the 99%+ of careful human transcription — but at 20–100× lower cost than human services, the trade is worth making for most work. We publish ranges rather than a single marketing number because a single number would not survive contact with your audio.

Methodology: Word Error Rate is (substitutions + insertions + deletions) / total words — the industry-standard formula. Ranges above come from in-house testing on standardized audio, human-verified. See our editorial standards for full testing methodology.

Accuracy by language

The language is detected from the audio — you do not select it before uploading. Accuracy is not uniform across all 99, and pretending otherwise would be useless to anyone deciding whether this will work for their recording. Here is roughly where each group lands on clear audio.

Tier 1

Around 5% word error rate

EnglishSpanishFrenchGermanItalianPortugueseDutchPolishRussianJapanese
Tier 2

Around 8–12% word error rate

ArabicChineseKoreanHindiTurkishVietnameseThaiIndonesianHebrewCzechSwedishNorwegianDanishFinnishGreekUkrainian

Plus 73 more languages

Including Welsh, Swahili, Filipino, Bengali, Punjabi, Tamil, Telugu, Marathi, Urdu, Persian, Romanian, Hungarian, Bulgarian, and Croatian. Accuracy varies more widely here, and depends on audio quality at least as much as on the language itself.

These figures come from in-house testing on standardized audio, verified against human transcripts. They describe clear recordings with a single speaker; a noisy multi-speaker file in a Tier 1 language can easily perform worse than a clean recording in Tier 2. Audio quality moves accuracy further than language choice does.

How to get a more accurate transcript

Accuracy is decided mostly at recording time, not at transcription time. The difference between a 3% and a 15% error rate is usually the recording, not the model — and every one of these is free to fix.

1. Get the microphone closer

The single largest factor, and it costs nothing. A phone lying on a table two metres away picks up as much room as speech; the same phone at arm's length is a different recording. If you can only change one thing, change this.

2. Record each speaker separately when you can

Multi-track recorders and most conferencing tools can save a file per participant. Uploading those separately gives near-perfect speaker separation, because the separation is in the recording rather than inferred from it.

3. Leave a beat between speakers

Overlapping speech is where both transcription and speaker labelling degrade fastest. In an interview, a half-second gap before responding is worth more than a better microphone.

4. Prefer a quiet room over noise reduction

Aggressive noise reduction removes parts of the speech signal along with the noise. A slightly noisy raw recording usually transcribes better than a heavily processed one.

5. Skip music beds under speech

Music shares frequency range with the voice and has no predictable structure to separate out. If a recording will be transcribed, add the music afterwards.

If a recording already exists and cannot be redone, upload it anyway and check the result against the list below — a transcript at 90% is still far faster to correct than typing from scratch, provided you know where to look.

Before you publish: what to check

Automatic transcription fails in predictable places, which makes reviewing it much faster than reading every word. These five account for most of what goes wrong, roughly in the order they are worth checking.

1. Proper names and brand names

The most common single error, and the most visible one. A name the model has not seen becomes the nearest common word. Search the transcript for each name once and fix every instance at the same time.

2. Numbers, dates, and currency

Spoken numbers are ambiguous in ways written ones are not — "fifteen" and "fifty" differ by one unstressed syllable, and a misheard decimal changes a figure by an order of magnitude. Check anything you plan to quote.

3. Technical and domain jargon

Field-specific vocabulary and acronyms sit outside everyday language. Medical, legal, and engineering recordings need the heaviest pass here.

4. Homophones in context

Their, there, they're — and worse in a domain context, where principal and principle change the meaning of a sentence entirely. These read fluently while being wrong, so they survive a skim.

5. Attribution across interruptions

When two people talk over each other, the words are usually right but the speaker label can land on the wrong turn. Check any passage where the conversation gets fast before quoting anyone.

For anything quoted publicly — journalism, research, legal work — play the audio back at the timestamp of each quote before it goes out. The editor makes this quick: click a word and playback jumps to that moment.

Which export format to choose

One processing run produces all six. They differ in what timing they carry, which is usually what decides the choice.

FormatUse it forTiming it carriesOpens in
TXTReading, pasting into a doc, feeding another toolLine timestamps, optionalAnything
DOCXHanding to an editor or client for tracked changesLine timestamps and speaker labelsWord, Google Docs, Pages
SRTSubtitles for YouTube, Vimeo, Premiere, ResolveCues, capped at 80 characters and 6 secondsEvery video editor and player
VTTSubtitles for HTML5 video and HLS streamsCues, same rules as SRTBrowsers via the track element
CSVAnalysis — counting, filtering, pivoting by speakerOne row per segment, with start and endExcel, Sheets, pandas
JSONDeveloper pipelines, research coding, custom playersPer-word start, end, and confidenceAny language with a JSON parser

The one worth knowing about is JSON. It is the only export carrying a start time, end time, and confidence score for every individual word, which is what makes it the right input for a custom player, a research coding tool, or any pipeline that needs to line text up against audio precisely.

Long recordings, and a lot of them

A one-hour file takes about 5–10 minutes. What moves that number is audio quality and current load rather than length — processing is roughly proportional, so a three-hour recording takes about three times as long, not longer than that.

File limits in practice

200 MB without an account, 5 GB with one, up to 10 hours per file. For reference, an hour of typical MP3 audio is roughly 30–60 MB, so the free preview comfortably accepts a long recording — it transcribes the first five minutes of it.

Uncompressed audio

WAV and AIFF are around ten times larger than MP3 for the same audio, which is what usually pushes a file over a cap. Converting to MP3 or M4A before uploading does not measurably change accuracy at normal bitrates.

Transcribing a back catalogue

Bulk upload takes a batch of files in one go and processes them together, and bulk export downloads the finished transcripts as a set rather than one at a time. For an archive — a podcast back catalogue, a season of interviews, a term of lectures — that is the difference between an afternoon of clicking and a single operation.

Each file keeps its own speaker labels and timestamps, and every export format is available across the batch.

How speaker labels work, and where they slip

Speakers are separated automatically and labelled Speaker 1, Speaker 2, and so on, in the order they first talk. Rename one in the editor and the change applies to every line that speaker has — you are not editing labels one at a time.

Where it works well

Two to four speakers who take turns, recorded close to the microphone, in a room without much echo. A standard interview or a small meeting is close to the best case, and separation there is reliable enough to quote from directly.

Where it slips

  • Crosstalk — when two people speak at once, the words usually survive but the label can land on the wrong turn.
  • Similar voices — speakers of the same register and accent are harder to tell apart than distinct ones.
  • Phone and compressed audio — narrowband recordings strip out much of what distinguishes one voice from another.
  • Large groups — the more people in the room, the more the labels drift, particularly with brief interjections.

The reliable fix is to record each speaker on a separate track where the setup allows it, then upload those separately — separation is then part of the recording rather than inferred from it. Failing that, check attribution at any point where the conversation speeds up.

How VexaScribe compares

The tools most people also evaluate when they search for audio-to-text. Verified July 2026.

FeatureVexaScribeHappyScribeOtterRev (AI)
Free tier30 min at signup10 min AI trial300 min/mo45 min/mo
Max file size5 GBNo hard limit stated4 hr/fileStandard file caps
Languages99 (auto-detect)60+ free / 150+ paid515
Speaker labelsEvery plan freeIncludedIncludedIncluded
Effective cost/hr$0.20–$0.60~$5–$17~$3.40 (Pro)$15 ($0.25/min)

Sources: HappyScribe pricing, Otter pricing, Rev pricing, VexaScribe pricing (all verified July 28, 2026). Effective cost per hour is each vendor's cheapest paid plan divided by included minutes.

Honest picks: Otter is the better choice if you mainly need live meeting captions rather than uploaded-file transcription. Rev is better when you want a certified-human upgrade path on the same platform. For long files, 99 languages, developer exports, and lowest cost per hour, VexaScribe wins on our own comparison. For a broader ranking see our best transcription software 2026 analysis.

What happens to your audio

Plenty of recordings are confidential — client interviews, medical notes, legal calls, unreleased material. It is a fair thing to ask before uploading, and the answer should be specific rather than reassuring.

In transit and at rest

Uploads go over TLS 1.2 or higher, and files are stored encrypted at rest.

Training

We do not train AI models on your audio or your transcripts, and we do not sell user data.

Deletion

You can delete any file from the dashboard at any time, and deleting your account is self-serve rather than a support request.

The anonymous preview

A preview run without an account is not attached to any profile. Only the first five minutes are transcribed, and the upload is truncated server-side before processing rather than after.

If your work carries a specific compliance obligation — HIPAA, legal privilege, a client confidentiality clause — check the terms against it rather than taking a summary on trust. That applies to any tool, not only this one.

Pricing

A 5-minute preview with no account, 30 minutes free when you sign up, no card either way. Paid plans are pay-for-what-you-use with no per-seat fees.

Starter

$2/month

200 min/month

Basic

$5/month

1,000 min/month

Pro

$10/month

2,500 min/month

Studio

$20/month

6,000 min/month

Cancel anytime. See all plans, including Team.

Frequently asked questions

How do I transcribe audio to text?

Upload an audio or video file, wait for it to process, then edit or export the transcript. A 60-minute recording typically completes in 5-10 minutes. Speaker labels and timestamps are added automatically — you don't configure anything first.

Is it free?

There are three numbers. A 5-minute preview with no account and no card. 30 minutes free when you sign up, still no card. After that, plans start at $2/month for 200 minutes. Nothing expires into a trial you forget to cancel.

Do transcripts include timestamps, and do exports keep them?

Yes to both. Segment-level timestamps appear at each speaker turn on screen and in the TXT and DOCX exports. SRT and VTT carry subtitle cues. The JSON export, available with an account, carries word-level timings for every individual word.

Can I get a timestamped transcript without paying?

Yes. The free 5-minute preview includes timestamps and speaker labels, and you can download it as TXT or SRT without creating an account.

How accurate is it?

Around 95% on clear English audio with a single speaker. Real-world accuracy varies: clear podcast audio runs 3-6% word error rate, noisy interviews 8-15%, and strong accents or heavy jargon 10-20%. Review transcripts before publishing critical content.

What formats and file sizes can I upload?

Ten audio formats (MP3, WAV, M4A, FLAC, OGG, AAC, AIFF, WMA, AMR, OPUS) and seven video formats (MP4, MOV, AVI, MKV, WebM, FLV, WMV). Free preview: up to 200 MB. With a free account: up to 5 GB and 10 hours per file.

Does it identify different speakers?

Yes, automatically, and it works in the free preview. Speakers are labeled Speaker 1, Speaker 2, and so on. You can rename them in the editor — to "Host", "Guest", or real names — and the change applies across the whole transcript.

Can I translate the transcript?

Yes, in place, without re-uploading. Free translations are included with an account. Premium translation handles long recordings better — it keeps terminology and names consistent from start to finish rather than drifting partway through.

What languages are supported?

99 languages including English, Spanish, French, German, Italian, Portuguese, Dutch, Russian, Polish, Japanese, Chinese, Korean, Arabic, Turkish, and Hindi. The language is detected automatically — you don't select it before uploading.

Can I transcribe video?

Yes. Upload MP4, MOV, AVI, MKV, WebM, FLV, or WMV and the audio track is extracted automatically. Export as SRT or VTT for subtitles, or as text to repurpose the content.

How long does a 1-hour file take?

Usually 5-10 minutes, depending on audio quality and current load. You can close the tab — with an account, the transcript waits in your dashboard when it's done.

Is my audio private?

Audio transits over TLS 1.2+ and is stored encrypted at rest. We do not train AI models on your audio and we do not sell user data. You can delete files at any time from your dashboard, and account deletion is self-serve.

Start transcribing

30 minutes free when you sign up, no credit card. Upload any audio file and see the result yourself.