Settings
Choosing it helps with short or noisy clips. Detection listens to the first 30 seconds.
In the transcript, when copying and in the .txt file. Subtitles always have them.
On WebAssembly, Base took about 25 seconds per minute of audio on a 2012 desktop processor, and recent laptops are faster. Tiny takes about half that and Small about two and a half times more. Keep this tab open during long files.
Whisper Base (80 MB) is downloaded once from Hugging Face and then kept by your browser. Your audio never leaves this device.
Transcript
Transcription runs in your browser: the audio and the text are never uploaded. Your settings and the last transcript are kept on this device only, until you clear them.
Embed this tool on your site
Free and without an account: copy the code and paste it into your website, blog or docs.
About the Audio to Text Transcriber
Turn interviews, lectures, voice notes and videos into text without uploading them anywhere. The transcriber runs OpenAI's Whisper speech recognition model inside your browser: drop a file or record from the microphone, choose a model and a language or let it be detected, and the text appears as it is transcribed, with timestamps you can click to hear each line. Export it as plain text or as SRT and VTT subtitles.
How to use the Audio to Text Transcriber
-
Add the audio
Drop an audio or video file on the page (MP3, WAV, M4A, OGG, WebM, MP4, MOV…) or press Record from the microphone. The file is decoded in your browser, and a player lets you listen to it.
-
Choose the model and the language
Base is the recommended model; Tiny is faster and Small more accurate. Keep Detect automatically or pick the language, and choose Translate to English if you want the text in English. Press Transcribe: the first time, the model is downloaded once and kept by your browser.
-
Review and export
The transcript grows as each 30-second stretch of audio is processed. Click a timestamp or a sentence to play the audio from there, then copy the text or download it as TXT, SRT or VTT.
When to use it
Interviews and meetings
Get a searchable transcript of an interview, a podcast or a recorded meeting to quote, summarize or take notes from, without sending a confidential recording to a transcription service.
Subtitles for videos
Generate an SRT or VTT file for a YouTube video, an online course or a social media clip, then fix any misheard words in a subtitle editor before publishing.
Voice notes and lectures
Turn voice messages, dictated ideas or class recordings into text you can read, search and copy, in the language they were recorded in or translated into English.
Frequently asked questions
Is my audio uploaded anywhere?
No. The audio is decoded and transcribed by your browser, on your device, and the text stays there too. The only downloads are the Transformers.js library and the Whisper model files, fetched from jsDelivr and Hugging Face the first time; your recording is never part of any request. Your settings and the last transcript are saved in this browser only, until you clear them.
Which Whisper model should I choose, and how big is it?
Whisper is OpenAI's open speech recognition model, released under the MIT license; this tool uses the ONNX versions published on Hugging Face. Tiny (39 million parameters, about 44 MB) is the fastest and makes the most mistakes. Base (74 million, about 80 MB) is the recommended balance. Small (244 million, about 250 MB) is the most accurate, but about two and a half times slower than Base. With WebGPU larger files are used (about 120, 210 and 590 MB), because they run faster on a graphics card. Each model is downloaded once; after that it loads from the browser cache and also works offline.
How long does a transcription take?
It depends on your device and the model. With WebGPU, available in recent versions of Chrome and Edge, a minute of audio usually takes a few seconds. Without it the model runs on the processor with WebAssembly: on a 2012 desktop processor, Base needed about 25 seconds per minute of audio, Tiny about half that and Small about two and a half times more, and recent laptops are faster. Long recordings take a while, so keep the tab open; you can read and copy the text while it grows.
Which file formats and languages does it support?
Any audio or video file your browser can play: usually MP3, WAV, M4A and AAC, OGG and Opus, FLAC, WebM and MP4, and MOV in most browsers. Whisper understands about 100 languages and detects the language from the first 30 seconds; it is most accurate in widely spoken languages such as English, Spanish, Portuguese, French or German. It can also translate speech in any of them into English.
How accurate is it, and what are its limits?
Whisper is accurate on clear speech, but no transcript is perfect: names, technical terms, overlapping voices, music and background noise cause mistakes, and the smaller models make more of them. It does not tell speakers apart, and over long silences or music it can occasionally invent a short phrase. Timestamps mark each segment, a sentence or two, not each word, so review subtitles before you publish them.
Related tools
Browse all free online tools on Tooly — every one runs in your browser, with no account required.