Transcribe Audio and Video to Text
How audio to text transcription works on this page
Turn speech from MP3, WAV, FLAC, M4A, AAC, OGG, Opus, WebM or supported video files into editable text. Choose automatic language detection or a known source language, optionally translate the transcript, and add timestamps or speaker labels for meetings, interviews and calls. The result can be copied or downloaded as a text file, with DOCX available when the processing result provides it. Speaker separation and wording depend on the recording quality and selected model, so important transcripts should still be reviewed.
Cost: Free within the available allowance. Usage limits apply; see the guide below.
How to transcribe audio or video to text
Add one recording
Click Upload an audio file to transcribe or drop the file onto the preview area. Audio and video files of up to 200 MB work. Files you uploaded or created in other tools are listed under Files.
Pick the AI model
Keep Gemini 3.5 Flash Lite · fast unless you have a reason to change it. The model section below helps you choose.
Set the language
Leave Audio Language on Auto-detect or choose the spoken language. Choose a target under Translate to (optional) only if you want the text in another language.
Choose the transcript detail
Switch on Include timestamps, Identify different speakers or both for interviews, meetings and calls.
Transcribe and export
Click Transcribe Audio and keep the tab open until the text appears under Transcribed Text. Then use Copy to Clipboard, Download as TXT or DOCX.
Settings and when to change them
- AI Model
- Gemini 3.5 Flash Lite · fast is the default. Gemini 3.6 Flash · advanced is the larger Google model. Both follow instructions, so they handle translation, timestamps and speaker labels. Experimental is Z.ai's GLM-ASR speech recognition model for clips of up to 30 seconds. The next section helps you choose.
- Audio Language
- Default Auto-detect. The list offers English, Spanish, French, German, Italian, Portuguese, Russian, Japanese, Chinese and Arabic. Setting it helps with short clips, strong accents and speech full of names or loanwords. If your language isn't listed, keep Auto-detect and the model identifies the language itself.
- Translate to (optional)
- Default No translation (original language). To translate audio to text in another language, pick English, Spanish, French, German, Italian, Portuguese, Russian, Japanese, Chinese, Arabic, Hindi, Korean, Dutch, Polish or Turkish. The model transcribes and translates in one pass and returns only the translation, with speaker labels kept. If you need both versions, run the job twice.
- Include timestamps
- Off by default. When on, markers such as
[00:01:23](hours:minutes:seconds) appear at regular intervals, each on a new line. The model writes them itself, so use them to find passages, not as exact timing. For timed subtitle files, use Video to Subtitles. When off, the transcript is plain text split into readable paragraphs. - Identify different speakers
- Off by default. Adds
Speaker 1:,Speaker 2:and so on, starting a new line at each change. With timestamps, a line reads[00:04:10] Speaker 2: …. Speakers are told apart by voice only: the audio is mixed down to one channel first, so recording each person on a separate stereo channel doesn't help. Replace the numbered labels with real names by using find and replace in your editor.
Which model to choose
Your AI Model choice is used for your recording.
Gemini 3.5 Flash Lite · fast
The default. Quick and reliable for clear speech such as voice memos, lectures and single-speaker recordings, and a good fit for long files.
Gemini 3.6 Flash · advanced
As a rule of thumb, pick it for difficult audio such as crosstalk, strong accents or technical vocabulary.
Experimental: 30 seconds maximum
Z.ai's GLM-ASR model for short clips; longer clips are rejected with an error when you click Transcribe Audio. It is built for plain speech recognition, so choose a Gemini model for translation, timestamps and speaker labels.
What happens to your audio, and how to get cleaner text
Before transcription, FFmpeg on our server takes the first audio track and prepares mono audio at 16 kHz; the Gemini route uses FLAC. Video frames, cover art and metadata tags are dropped. Only the prepared audio goes to the selected AI service, and the processing copy is deleted from our server after the request.
Every upload is converted to mono audio at 16 kHz before transcription. An MP3 or M4A can be smaller than a stereo WAV, but prior lossy compression changes the recording and is not guaranteed to produce an identical transcript. Keep the speech clear when choosing a smaller export.
That helps with the 200 MB limit. CD-quality stereo WAV takes about 10 MB per minute, so roughly 19 minutes fit. A 128 kbps MP3 takes under 1 MB per minute, so about 3.5 hours fit.
Formats and limits
- Input formats
- Audio: MP3, WAV, FLAC, M4A, AAC, OGG, Opus, WebM. Video: MP4, M4V, MOV, MKV, WebM, AVI, MPEG/MPG, OGV. The file must contain an audio track.
- Output formats
- Text on the page, TXT (UTF-8) and DOCX
- Maximum file size
- 200 MB per file
- Recording length
- Experimental: 30 seconds. Gemini models: no fixed length, but the job must finish within 60 minutes and very long transcripts can hit the model's output limit.
- Files per job
- 1
- Processing
- On our servers. FFmpeg prepares the audio, then an AI model transcribes it. Jobs can run for up to 60 minutes.
- External AI services
- Gemini models: the prepared audio goes to Google's Gemini models, through Google's Gemini API or through OpenRouter. Experimental: the audio goes to Z.ai's GLM-ASR API.
- Usage limits
- Daily usage limits apply. The tool shows an error when you reach one.
- Watermark
- None
- Storage
- Uploads and results are deleted automatically after 48 hours for guests and after 30 days for free accounts. Paid plans can keep files longer.
Frequently asked questions
Can I get a transcript from a video file?
Yes. Upload MP4, M4V, MOV, MKV, WebM, AVI, MPEG or OGV. Only the first audio track is transcribed, so if a video has several language tracks and yours isn't first, extract it in a video editor beforehand. A video without sound fails with an error saying it has no audio track.
Is my recording sent to an AI service?
Yes, the audio is. With the Gemini options, the mono 16 kHz audio goes to Google's Gemini models, either directly through Google's Gemini API or through OpenRouter. With Experimental, it goes to Z.ai. The picture and the file's metadata are not sent. Your upload and the DOCX are deleted automatically after 48 hours for guests and after 30 days for free accounts.
Why does the transcript stop before the end of the recording?
AI models can only return a limited amount of text per request. When that limit is reached, a warning above the text says the transcription was truncated. Cut the recording into parts with Split Audio, which delivers them as a ZIP, and transcribe the parts one by one.
What does "No text could be transcribed from this audio" mean?
The model returned no text. This happens with silence, music or noise without clear speech, and with speech that is very quiet. Play the file to check that the track actually contains speech, then try again with Audio Language set to the spoken language.
My file is larger than 200 MB. What can I do?
MP3 or M4A can reduce file size compared with WAV. The tool converts every input to 16 kHz mono, but prior lossy compression can affect recognition, so check that speech remains clear in the smaller recording. For a large video, Compress Video accepts bigger files. Keep its audio track, then pick the smaller result under Files here.
How accurate is the transcript?
It depends mostly on the recording. One voice close to the microphone usually needs only light editing; background music, echo, phone-line audio and people talking over each other cause most errors. Check names, numbers and technical terms first. The text may also leave out filler words and false starts, so it isn't a strict word-for-word record.