free transcribe
Examples

What the output looks like, on audio you can hear.

Each sample: the recording, the transcript from a real browser run, a reference, and the export files. Nothing is typed in by hand.

Clear speech

Clean speech, one synthetic voice

A scripted passage read by the Windows text-to-speech voice. Evenly paced, no background noise, no overlapping speakers. This is the easiest kind of audio a speech model will ever hear, so treat the result as a best case, not a typical one.

Length
00:01:11
File
MP3, mono, 22.05 kHz, 48 kbps (encoded from a 16-bit WAV) · download the audio
Source
Generated by us in September 2026 with the Windows System.Speech text-to-speech engine from a passage we wrote. It is the smoke-test file used during development.
Licence
Our own material. You may download and reuse it.

FreeTranscribe output

00:00:00Thanks for making the time.

00:00:02I want to start with the part you said was messy.

00:00:06Right, so the messy part is that the file never arrives the way you expect it.

00:00:12Someone records on a phone, someone else on a laptop, and the levels are nothing alike.

00:00:19We stopped sending the audio anywhere.

00:00:22Once the whole thing ran on the machine that recorded it, half the problems were just gone.

00:00:29What surprised me was how much of the time went into the download, not the transcription.

00:00:35The model is about 200 megabytes, and after the first run it is cached.

00:00:41On the laptop an hour of audio takes about three minutes. On the desktop with a proper graphics card, well under one. That was the moment it stopped feeling like a demo.

00:00:54Let me add a pause here. And then continue with a second paragraph about subtitle formats.

00:01:01SRT works on every video platform.

00:01:05VTT is what browsers use.

00:01:08Plain text is for notes.

Download: TXT · SRT · VTT

Reference transcript reconstructed, not yet checked

Thanks for making the time. I want to start with the part you said was messy. Right, so the messy part is that the file never arrives the way you expect it. Someone records on a phone, someone else on a laptop, and the levels are nothing alike. We stopped sending the audio anywhere. Once the whole thing ran on the machine that recorded it, half the problems were just gone. What surprised me was how much of the time went into the download, not the transcription. The model is about 200 megabytes, and after the first run it is cached. On the laptop an hour of audio takes about three minutes. On the desktop with a proper graphics card, well under one. That was the moment it stopped feeling like a demo. Let me add a pause here. And then continue with a second paragraph about subtitle formats. SRT works on every video platform. VTT is what browsers use. Plain text is for notes.

The passage typed into the text-to-speech engine was not kept. This reference was reconstructed on 2026-09-25 from the generation notes and a second run of the same model in Node.js, then read for sense. Numbers are spelt out the way the browser runs produced them; a Node.js run of the same model wrote them as digits, and the script itself is not available to settle it. It still needs a listen-through by a person before it counts as checked.

Differences

thanks for making the time i want to start with the part you said was messy right so the messy part is that the file never arrives the way you expect it someone records on a phone someone else on a laptop and the levels are nothing alike we stopped sending the audio anywhere once the whole thing ran on the machine that recorded it half the problems were just gone what surprised me was how much of the time went into the download not the transcription the model is about 200 megabytes and after the first run it is cached on the laptop an hour of audio takes about three minutes on the desktop with a proper graphics card well under one that was the moment it stopped feeling like a demo let me add a pause here and then continue with a second paragraph about subtitle formats srt works on every video platform vtt is what browsers use plain text is for notes

Word error rate 0.0% against the reference (0 substituted, 0 inserted, 0 deleted, 165 reference words). Case and punctuation are ignored; numerals and spelt-out numbers are not merged.

Processing details

Run date
2026-09-25
Browser
Chrome 153.0.8010.48 (desktop), driven by scripts/record-sample-run.mjs
OS
Windows 11 Home 24H2 (build 26200)
CPU
Intel Core i7-11390H, 4 cores / 8 threads
GPU
Intel Iris Xe integrated graphics (WebGPU adapter reported as "intel · gen-12lp"); the machine also has an NVIDIA RTX 3050 Laptop GPU that Chrome did not pick
RAM
16 GB
Model
whisper-base.en, fp32 encoder + q4 decoder (about 206 MB)
Backend
WebGPU (graphics card)
Model cached
yes, from an earlier run
Time
25 seconds for 00:01:11 of audio
Real-time factor
0.36 (processing time divided by audio length; under 1 is faster than the recording)

Second run in the same browser profile, so the model came from the browser cache. The first run in a fresh profile, which downloaded the model from the Hugging Face Hub during the run, took 27.0 seconds on this connection. Time is the tool's own measurement from the start of the job to the result; the wall clock from file selection to result view was 25.5 seconds. Segment text is exactly as the tool emitted it, including where it started a sentence in lower case.

Background noise

Background noise, about 10 dB signal-to-noise

The clean clip with pink noise mixed in with ffmpeg. Measured over the whole file the speech sits about 10 dB above the noise, which is a busy room or a fan next to the microphone. The noise is synthetic and steady; real rooms have doors and other voices.

Length
00:01:11
File
MP3, mono, 22.05 kHz, 48 kbps · download the audio
Source
The clean-speech clip with ffmpeg's pink noise generator mixed in (anoisesrc, pink, -8 dB gain); speech RMS -18.5 dBFS, noise RMS about -28 dBFS.
Licence
Our own material. You may download and reuse it.

FreeTranscribe output

00:00:00Thanks for making the time.

00:00:02I want to start with the part you said was messy.

00:00:06Right, so the messy part is that the file never arrives the way you expect it.

00:00:12Someone records on a phone, someone else on a laptop, and the levels are nothing alike.

00:00:19We stopped sending the audio anywhere.

00:00:22Once the whole thing ran on the machine that recorded it, half the problems were just gone. What surprised me was how much of the time went into the download, not the transcription. The model is about 200 megabytes, and after the first run it is cached.

00:00:41On the laptop an hour of audio takes about three minutes.

00:00:44On the desktop with a proper graphics card well under one that was the moment it stopped feeling like a demo let me add a pause here and then continue with the second paragraph about subtitle formats SRT works on every video platform.

00:01:05VTT is what browsers use.

00:01:08Play text is for notes.

Download: TXT · SRT · VTT

Reference transcript reconstructed, not yet checked

Thanks for making the time. I want to start with the part you said was messy. Right, so the messy part is that the file never arrives the way you expect it. Someone records on a phone, someone else on a laptop, and the levels are nothing alike. We stopped sending the audio anywhere. Once the whole thing ran on the machine that recorded it, half the problems were just gone. What surprised me was how much of the time went into the download, not the transcription. The model is about 200 megabytes, and after the first run it is cached. On the laptop an hour of audio takes about three minutes. On the desktop with a proper graphics card, well under one. That was the moment it stopped feeling like a demo. Let me add a pause here. And then continue with a second paragraph about subtitle formats. SRT works on every video platform. VTT is what browsers use. Plain text is for notes.

Same passage and same caveat as the clean-speech sample: reconstructed, not yet checked by listening.

Differences

thanks for making the time i want to start with the part you said was messy right so the messy part is that the file never arrives the way you expect it someone records on a phone someone else on a laptop and the levels are nothing alike we stopped sending the audio anywhere once the whole thing ran on the machine that recorded it half the problems were just gone what surprised me was how much of the time went into the download not the transcription the model is about 200 megabytes and after the first run it is cached on the laptop an hour of audio takes about three minutes on the desktop with a proper graphics card well under one that was the moment it stopped feeling like a demo let me add a pause here and then continue with a the second paragraph about subtitle formats srt works on every video platform vtt is what browsers use plain play text is for notes

Word error rate 1.2% against the reference (2 substituted, 0 inserted, 0 deleted, 165 reference words). Case and punctuation are ignored; numerals and spelt-out numbers are not merged.

Processing details

Run date
2026-09-25
Browser
Chrome 153.0.8010.48 (desktop), driven by scripts/record-sample-run.mjs
OS
Windows 11 Home 24H2 (build 26200)
CPU
Intel Core i7-11390H, 4 cores / 8 threads
GPU
Intel Iris Xe integrated graphics (WebGPU adapter reported as "intel · gen-12lp")
RAM
16 GB
Model
whisper-base.en, fp32 encoder + q4 decoder (about 206 MB)
Backend
WebGPU (graphics card)
Model cached
yes, from an earlier run
Time
25 seconds for 00:01:11 of audio
Real-time factor
0.35 (processing time divided by audio length; under 1 is faster than the recording)

Model already cached from the clean-speech run in the same browser profile. Segment text is exactly as the tool emitted it.

Quiet recording

Quiet speech, 24 dB below the original

The clean clip turned down by 24 dB, so its loudest peak is at -25 dBFS. This is a recorder held too far from the speaker or a gain set too low. No noise was added; a quiet real recording would also have more hiss relative to the voice.

Length
00:01:11
File
MP3, mono, 22.05 kHz, 48 kbps · download the audio
Source
The clean-speech clip with ffmpeg volume=-24dB.
Licence
Our own material. You may download and reuse it.

FreeTranscribe output

00:00:00Thanks for making the time.

00:00:02I want to start with the part you said was messy.

00:00:06Right, so the messy part is that the file never arrives the way you expect it.

00:00:12Someone records on a phone, someone else on a laptop, and the levels are nothing alike.

00:00:19We stopped sending the audio anywhere.

00:00:22Once the whole thing ran on the machine that recorded it, half the problems were just gone.

00:00:29What surprised me was how much of the time went into the download, not the transcription.

00:00:35The model is about 200 megabytes, and after the first run it is cached.

00:00:41On the laptop an hour of audio takes about three minutes. On the desktop with a proper graphics card, well under one. That was the moment it stopped feeling like a demo.

00:00:54Let me add a pause here. And then continue with a second paragraph about subtitle formats.

00:01:01SRT works on every video platform.

00:01:05VTT is what browsers use.

00:01:08Plain text is for notes.

Download: TXT · SRT · VTT

Reference transcript reconstructed, not yet checked

Thanks for making the time. I want to start with the part you said was messy. Right, so the messy part is that the file never arrives the way you expect it. Someone records on a phone, someone else on a laptop, and the levels are nothing alike. We stopped sending the audio anywhere. Once the whole thing ran on the machine that recorded it, half the problems were just gone. What surprised me was how much of the time went into the download, not the transcription. The model is about 200 megabytes, and after the first run it is cached. On the laptop an hour of audio takes about three minutes. On the desktop with a proper graphics card, well under one. That was the moment it stopped feeling like a demo. Let me add a pause here. And then continue with a second paragraph about subtitle formats. SRT works on every video platform. VTT is what browsers use. Plain text is for notes.

Same passage and same caveat as the clean-speech sample: reconstructed, not yet checked by listening.

Differences

thanks for making the time i want to start with the part you said was messy right so the messy part is that the file never arrives the way you expect it someone records on a phone someone else on a laptop and the levels are nothing alike we stopped sending the audio anywhere once the whole thing ran on the machine that recorded it half the problems were just gone what surprised me was how much of the time went into the download not the transcription the model is about 200 megabytes and after the first run it is cached on the laptop an hour of audio takes about three minutes on the desktop with a proper graphics card well under one that was the moment it stopped feeling like a demo let me add a pause here and then continue with a second paragraph about subtitle formats srt works on every video platform vtt is what browsers use plain text is for notes

Word error rate 0.0% against the reference (0 substituted, 0 inserted, 0 deleted, 165 reference words). Case and punctuation are ignored; numerals and spelt-out numbers are not merged.

Processing details

Run date
2026-09-25
Browser
Chrome 153.0.8010.48 (desktop), driven by scripts/record-sample-run.mjs
OS
Windows 11 Home 24H2 (build 26200)
CPU
Intel Core i7-11390H, 4 cores / 8 threads
GPU
Intel Iris Xe integrated graphics (WebGPU adapter reported as "intel · gen-12lp")
RAM
16 GB
Model
whisper-base.en, fp32 encoder + q4 decoder (about 206 MB)
Backend
WebGPU (graphics card)
Model cached
yes, from an earlier run
Time
25 seconds for 00:01:11 of audio
Real-time factor
0.35 (processing time divided by audio length; under 1 is faster than the recording)

Model already cached from the clean-speech run in the same browser profile. Segment text is exactly as the tool emitted it.

Fast speech

Fast speech, 1.5 times the original pace

The clean clip sped up by half again with pitch preserved, so 71 seconds of speech fit in 48. Roughly the pace of a fast talker or a podcast played at 1.5x. The speed-up is uniform, unlike a person, who slows down for the hard parts.

Length
00:00:47
File
MP3, mono, 22.05 kHz, 48 kbps · download the audio
Source
The clean-speech clip with ffmpeg atempo=1.5.
Licence
Our own material. You may download and reuse it.

FreeTranscribe output

00:00:00Thanks for making the time. I want to start with the part you said was messy.

00:00:03Right, so the messy part is that the file never arrives the way you expect it.

00:00:07Someone records on a phone, someone else on a laptop, and the levels are nothing alike.

00:00:12We stopped sending the audio anywhere. Once the whole thing ran on the machine that recorded it, half the problems were just gone. What surprised me was how much of the time went into the download, not the transcription. The model is about 200 megabytes, and after the first run it is cached.

00:00:27On the laptop an hour of audio takes about three minutes.

00:00:30On the desktop with a proper graphics card, well under one.

00:00:33That was the moment it stopped feeling like a demo.

00:00:36Let me add a pause here.

00:00:37And then continue with the second paragraph about subtitle formats.

00:00:41SRT works on every video platform.

00:00:43VTT is what browsers use.

00:00:45plain text is 4 notes.

Download: TXT · SRT · VTT

Reference transcript reconstructed, not yet checked

Thanks for making the time. I want to start with the part you said was messy. Right, so the messy part is that the file never arrives the way you expect it. Someone records on a phone, someone else on a laptop, and the levels are nothing alike. We stopped sending the audio anywhere. Once the whole thing ran on the machine that recorded it, half the problems were just gone. What surprised me was how much of the time went into the download, not the transcription. The model is about 200 megabytes, and after the first run it is cached. On the laptop an hour of audio takes about three minutes. On the desktop with a proper graphics card, well under one. That was the moment it stopped feeling like a demo. Let me add a pause here. And then continue with a second paragraph about subtitle formats. SRT works on every video platform. VTT is what browsers use. Plain text is for notes.

Same passage and same caveat as the clean-speech sample: reconstructed, not yet checked by listening.

Differences

thanks for making the time i want to start with the part you said was messy right so the messy part is that the file never arrives the way you expect it someone records on a phone someone else on a laptop and the levels are nothing alike we stopped sending the audio anywhere once the whole thing ran on the machine that recorded it half the problems were just gone what surprised me was how much of the time went into the download not the transcription the model is about 200 megabytes and after the first run it is cached on the laptop an hour of audio takes about three minutes on the desktop with a proper graphics card well under one that was the moment it stopped feeling like a demo let me add a pause here and then continue with a the second paragraph about subtitle formats srt works on every video platform vtt is what browsers use plain text is for 4 notes

Word error rate 1.2% against the reference (2 substituted, 0 inserted, 0 deleted, 165 reference words). Case and punctuation are ignored; numerals and spelt-out numbers are not merged.

Processing details

Run date
2026-09-25
Browser
Chrome 153.0.8010.48 (desktop), driven by scripts/record-sample-run.mjs
OS
Windows 11 Home 24H2 (build 26200)
CPU
Intel Core i7-11390H, 4 cores / 8 threads
GPU
Intel Iris Xe integrated graphics (WebGPU adapter reported as "intel · gen-12lp")
RAM
16 GB
Model
whisper-base.en, fp32 encoder + q4 decoder (about 206 MB)
Backend
WebGPU (graphics card)
Model cached
yes, from an earlier run
Time
22 seconds for 00:00:47 of audio
Real-time factor
0.46 (processing time divided by audio length; under 1 is faster than the recording)

Model already cached from the clean-speech run in the same browser profile. Segment text is exactly as the tool emitted it.

Interview

Two voices in turn, synthetic interview

A scripted interview read by two Windows text-to-speech voices taking turns, with no overlap and no room. It shows what a two-person recording looks like without speaker labels: one stream of text, and you tell the speakers apart from the sense. A real interview adds crosstalk and echo, which this does not test.

Length
00:01:02
File
MP3, mono, 22.05 kHz, 48 kbps · download the audio
Source
Generated by us on 2026-09-25 with the Windows System.Speech engine from a dialogue we wrote, alternating the Zira and David voices.
Licence
Our own material. You may download and reuse it.

FreeTranscribe output

00:00:00Thanks for making the time. Can you walk me through how the recordings reach you?

00:00:05Mostly by email, which is the first problem.

00:00:09A 40-minute interview at a decent bit rate is 30 megabytes, so people compress it, and then I get the compressed copy.

00:00:17And the compression is what hurts the transcript?

00:00:20Not as much as you would think.

00:00:23What hurts is when two people talk over each other, or when the microphone is on the table between them and the room has a lot of echo.

00:00:31So what would you tell someone recording their first interview?

00:00:35Put the recorder closer to the quieter person.

00:00:38Say both names at the start.

00:00:40And do not stop the recording to check it, because you will forget to start it again.

00:00:46That last one sounds like experience.

00:00:49Twice.

00:00:50The second time was with a professor whose flight left in an hour, so I had to reconstruct the middle third from memory.

00:00:57That was the day I stopped trusting myself and started trusting the red light.

Download: TXT · SRT · VTT

Reference transcript exact script

Thanks for making the time. Can you walk me through how the recordings reach you? Mostly by email, which is the first problem. A forty minute interview at a decent bitrate is thirty megabytes, so people compress it, and then I get the compressed copy. And the compression is what hurts the transcript? Not as much as you would think. What hurts is when two people talk over each other, or when the microphone is on the table between them and the room has a lot of echo. So what would you tell someone recording their first interview? Put the recorder closer to the quieter person. Say both names at the start. And do not stop the recording to check it, because you will forget to start it again. That last one sounds like experience. Twice. The second time was with a professor whose flight left in an hour, so I had to reconstruct the middle third from memory. That was the day I stopped trusting myself and started trusting the red light.

This is the exact text given to the text-to-speech engine, so it is the ground truth for what was said.

Differences

thanks for making the time can you walk me through how the recordings reach you mostly by email which is the first problem a forty 40 minute interview at a decent bit bitrate rate is thirty 30 megabytes so people compress it and then i get the compressed copy and the compression is what hurts the transcript not as much as you would think what hurts is when two people talk over each other or when the microphone is on the table between them and the room has a lot of echo so what would you tell someone recording their first interview put the recorder closer to the quieter person say both names at the start and do not stop the recording to check it because you will forget to start it again that last one sounds like experience twice the second time was with a professor whose flight left in an hour so i had to reconstruct the middle third from memory that was the day i stopped trusting myself and started trusting the red light

Word error rate 2.3% against the reference (3 substituted, 1 inserted, 0 deleted, 173 reference words). Case and punctuation are ignored; numerals and spelt-out numbers are not merged.

Processing details

Run date
2026-09-25
Browser
Chrome 153.0.8010.48 (desktop), driven by scripts/record-sample-run.mjs
OS
Windows 11 Home 24H2 (build 26200)
CPU
Intel Core i7-11390H, 4 cores / 8 threads
GPU
Intel Iris Xe integrated graphics (WebGPU adapter reported as "intel · gen-12lp")
RAM
16 GB
Model
whisper-base.en, fp32 encoder + q4 decoder (about 206 MB)
Backend
WebGPU (graphics card)
Model cached
yes, from an earlier run
Time
28 seconds for 00:01:02 of audio
Real-time factor
0.45 (processing time divided by audio length; under 1 is faster than the recording)

Model already cached. Two synthetic voices alternate with no overlap, which is easier than a real interview; the tool has no speaker labels, so the two voices come out as one stream. Segment text is exactly as the tool emitted it.

Another language

Dutch, one synthetic voice, multilingual model

A Dutch passage read by the Windows Dutch voice, transcribed with the multilingual base model and the language set to Dutch. This is the first run of that model in the tool; it is offered in the picker as untested for exactly this reason, and this row is the start of testing it.

Length
00:00:41
File
MP3, mono, 22.05 kHz, 48 kbps · download the audio
Source
Generated by us on 2026-09-25 with the Windows System.Speech engine (Microsoft Frank, nl-NL) from a passage we wrote.
Licence
Our own material. You may download and reuse it.

FreeTranscribe output

00:00:00bedankt dat je tijd hebt vrij gemaakt.

00:00:02Ik wil beginnen met het deel dat volgens jouw rommelig was.

00:00:07Het rommelige deel is dat het bestand nooit aankomt zoals je verwacht.

00:00:12De een neemt op met een telefoon, de ander met een laptop en de nivose lijken nergens op elkaar.

00:00:19We zijn gestopt met het versturen van audio.

00:00:22Zo draa alles draaiden op de machine die de opname maakte, was de helft van de problemen gewoon weg.

00:00:28Iedereen wilde het transcript in een ander formaat, en niemand wilde erop wachten.

00:00:34Platetext voor de notities, ondertitels voor de video mensen, en niets daar tussen.

Download: TXT · SRT · VTT

Reference transcript exact script

Bedankt dat je tijd hebt vrijgemaakt. Ik wil beginnen met het deel dat volgens jou rommelig was. Het rommelige deel is dat het bestand nooit aankomt zoals je verwacht. De een neemt op met een telefoon, de ander met een laptop, en de niveaus lijken nergens op elkaar. We zijn gestopt met het versturen van audio. Zodra alles draaide op de machine die de opname maakte, was de helft van de problemen gewoon weg. Iedereen wilde het transcript in een ander formaat, en niemand wilde erop wachten. Platte tekst voor de notities, ondertitels voor de videomensen, en niets daartussen.

This is the exact text given to the text-to-speech engine.

Differences

bedankt dat je tijd hebt vrij vrijgemaakt gemaakt ik wil beginnen met het deel dat volgens jou jouw rommelig was het rommelige deel is dat het bestand nooit aankomt zoals je verwacht de een neemt op met een telefoon de ander met een laptop en de niveaus nivose lijken nergens op elkaar we zijn gestopt met het versturen van audio zo zodra draa alles draaide draaiden op de machine die de opname maakte was de helft van de problemen gewoon weg iedereen wilde het transcript in een ander formaat en niemand wilde erop wachten platte tekst platetext voor de notities ondertitels voor de video videomensen mensen en niets daar daartussen tussen

Word error rate 13.1% against the reference (8 substituted, 4 inserted, 1 deleted, 99 reference words). Case and punctuation are ignored; numerals and spelt-out numbers are not merged.

Processing details

Run date
2026-09-25
Browser
Chrome 153.0.8010.48 (desktop), driven by scripts/record-sample-run.mjs
OS
Windows 11 Home 24H2 (build 26200)
CPU
Intel Core i7-11390H, 4 cores / 8 threads
GPU
Intel Iris Xe integrated graphics (WebGPU adapter reported as "intel · gen-12lp")
RAM
16 GB
Model
whisper-base (multilingual), fp32 encoder + q4 decoder (about 206 MB)
Backend
WebGPU (graphics card)
Model cached
no, downloaded during this run
Time
22 seconds for 00:00:41 of audio
Real-time factor
0.54 (processing time divided by audio length; under 1 is faster than the recording)

First run of the multilingual base model in this browser profile, so its download (about 206 MB) is included in the time. Language set to Dutch in the picker. Segment text is exactly as the tool emitted it.

Real recordings still to make

Everything above is synthetic speech, some of it processed with ffmpeg. That tests the model against a condition, not a recording made under it. These real recordings are still to make, each needing audio we may publish.

CategoryWhat it will beStatus
Background noiseA real room with a real cheap microphone, rather than synthetic noise mixed in.Planned
AccentA non-native English speaker, recorded with consent.Planned
InterviewTwo real people, one microphone, some overlap.Planned
LectureTwenty minutes of a real lecture, with permission.Planned
Difficult namesA passage dense with surnames and place names.Planned
Technical vocabularySoftware and medical vocabulary read by a person.Planned

How a sample gets onto this page

The file is dropped on the transcriber in a named browser on a named machine. The segments are copied from that run unchanged. The reference is the script given to the speech engine, or a person's listen-through; until then it is marked reconstructed. Word error rate is computed by the code that renders this page, so the number and the marked differences always agree.

To check it yourself, drop the same file on the transcriber.