Can I extract the text from a video?

Yes. Choose or drop the MP4, MOV, WEBM or MKV; its sound is read on your device and transcribed the same way as an audio file, with speaker labels and estimated timestamps, and you can download it as text or as subtitles (SRT or VTT). The times are estimates within each four-minute piece, so the subtitles are a rough guide rather than word-accurate.

What languages are supported?

Choose the spoken language from the list — the fifteen languages this site is written in — or leave it on automatic, which also suits a conversation that switches language mid-sentence, such as Arabic and English. The transcript keeps each part in the language it was spoken in.

Can I get subtitles or a spreadsheet?

Yes. The finished transcript exports to Word, PDF, plain text, subtitle files (SRT, VTT), and a CSV spreadsheet with one row per speaker turn and its estimated start and end time — pick whichever the next step needs.

Is my recording stored?

Our server stores neither the audio nor the text. The four-minute pieces of sound pass through Lovable’s AI gateway to OpenAI, and the text to Google’s Gemini; those services handle what they receive under their own terms. Our server notes your IP address for the hourly limit and deletes it after about two hours. A transcript is kept only if you save it: in this browser, or in your account when you are signed in.

Transcribe audio to text

Cloud step

Upload a recording — a meeting, an interview, a lecture, a voice note — and get back a clean transcript split by speaker, which you can export as Word, PDF, plain text, subtitles or a spreadsheet.

It is built for real conversations, including ones that switch language mid-sentence, and it splits the text by speaker — speakers inferred from the conversation, not identified by voice — rather than returning one undivided block.

Unlike the other tools on this site, transcription uses a server: this site does not run speech recognition on your device. The page asks before it sends anything. The recording goes in four-minute pieces, without its file name, through our server and Lovable’s AI gateway to OpenAI’s speech-recognition model, and the text to Google’s Gemini for tidying. Our server stores neither; those services follow their own terms.

How to transcribe audio to text

Choose or drop a recording or a video.
Choose the spoken language, or leave it on automatic for a conversation that mixes languages.
Wait while it is transcribed and split by speaker.
Export as Word, PDF, plain text, subtitles or a CSV spreadsheet with estimated times.

Supported formats and limits

Input: MP3, WAV, M4A, OGG, FLAC, OPUS, AMR; MP4, MOV, WEBM, MKV and other video (sound only)
Output: Word, PDF, TXT, Markdown, SRT or VTT subtitles, CSV spreadsheet, JSON
Languages: Automatic, including mixed conversations, or one of fifteen chosen from the list
Speakers: Labelled separately, inferred from the conversation rather than identified by voice
Times: Estimated within each four-minute piece
Hourly limit: 30 four-minute pieces (two hours of audio) per hour from one IP address
Where it runs: Our server function, then Lovable’s AI gateway to OpenAI (speech) and Google Gemini (text); our server stores neither, and those services follow their own terms

Not for: Perfect accuracy on poor recordings — heavy background noise and people talking over each other still cause mistakes.

Exactly what is sent, and where

Nothing is sent until you agree in the dialog that Transcribe opens, unless you chose not to be asked again. Your browser then decodes the recording, or the sound of a video, mixes it to one channel at 16 kHz and cuts it into four-minute pieces, each sent as a WAV file named part-1.wav, part-2.wav and so on: the sound, without your file name.

Each piece goes to our server function on Lovable Cloud, which passes it through Lovable’s AI gateway to OpenAI’s gpt-4o-transcribe speech model, with your custom vocabulary, the language if you chose one, and the last 1,200 characters already transcribed. The text of all pieces, with the speaker names and vocabulary you entered, goes the same way to Google’s Gemini model to be split into speakers and given a title, summary and chapters, and the finished transcript once more for key points and action items.

Our functions do not write the audio or the text to storage. They record your IP address with each request, for the hourly limit, and remove it once it is more than two hours old. What Lovable, OpenAI and Google keep is set by their own terms. Save to library keeps the transcript in this browser, or in your account if you are signed in; Share link uploads the exported file to our storage for 24 hours, as the page says beside the button.

How much can be transcribed in an hour

Our server accepts thirty pieces per hour from one IP address, which is two hours of audio. If a piece is refused or fails, the pieces already finished are kept, and Retry failed recordings carries on from the piece that failed rather than starting again.

How speakers, times and language are handled

Speakers are not recognised by their voices. Gemini infers them from the text, from questions, answers, greetings and names said aloud, and uses the expected speaker names you type for its labels. If that step fails, you still get the text, in four-minute blocks marked “Speaker not identified”.

Leave the language on automatic for a conversation that mixes languages, or choose the spoken language to help the speech model with a single-language recording.

Times are known only at the edges of each four-minute piece; turns inside a piece are spaced evenly and marked with ~. The SRT, VTT and CSV files carry these estimates, so they serve as a rough index into the recording rather than subtitles that follow the speech.

Frequently asked questions

Can I extract the text from a video?: Yes. Choose or drop the MP4, MOV, WEBM or MKV; its sound is read on your device and transcribed the same way as an audio file, with speaker labels and estimated timestamps, and you can download it as text or as subtitles (SRT or VTT). The times are estimates within each four-minute piece, so the subtitles are a rough guide rather than word-accurate.
What languages are supported?: Choose the spoken language from the list — the fifteen languages this site is written in — or leave it on automatic, which also suits a conversation that switches language mid-sentence, such as Arabic and English. The transcript keeps each part in the language it was spoken in.
Can I get subtitles or a spreadsheet?: Yes. The finished transcript exports to Word, PDF, plain text, subtitle files (SRT, VTT), and a CSV spreadsheet with one row per speaker turn and its estimated start and end time — pick whichever the next step needs.
Is my recording stored?: Our server stores neither the audio nor the text. The four-minute pieces of sound pass through Lovable’s AI gateway to OpenAI, and the text to Google’s Gemini; those services handle what they receive under their own terms. Our server notes your IP address for the hourly limit and deletes it after about two hours. A transcript is kept only if you save it: in this browser, or in your account when you are signed in.