Dictate with your microphone for live text, or pick an audio file and have it transcribed on your own device - the recording is never uploaded.
- To dictate: press Start listening and speak. Words appear as you talk, and you can edit the text in place afterwards.
- To transcribe a recording: scroll to the file-transcription panel below the dictation panel and choose an MP3, WAV, M4A, OGG, or WEBM file up to 15 minutes long.
- Press Load transcription model once (the size is shown before the download starts) - it is cached, so later visits skip the wait.
- Pick the spoken language or leave it on automatic, then start the transcription. The progress bar fills and the status counts "window X of N" as each 30-second window finishes.
- Copy the finished text or download it as a .txt file. Both modes keep your audio and text on this device.
Speech to Text
Convert a SPEECH input into a TEXT output - one step, in the browser.
Typical use: when your downstream platform accepts TEXT but not SPEECH.
There is one setting: the Language dropdown beside the Load transcription model button. Leave it on Auto-detect, or name the spoken language to keep the file transcript on that language from start to finish.
To compare this tool against desktop and server-based alternatives before committing to a workflow, see the speech-to-text alternatives guide.
The page stacks two panels and both stay on screen - there is no mode to switch. The dictation panel on top types what you say as you speak, handy for dictating a note, an email, or a quick draft. The file-transcription panel sits below it and handles a recording you already have: a meeting, a voice memo, a lecture, or an interview saved as .mp3, .wav, .m4a, .ogg, or .webm. Pick the file, click Transcribe file, and the transcript appears in the output box with copy and download buttons.
File transcription runs on your own device with the Whisper AI model, so the recording is never uploaded to a server. The model downloads once (about 40 MB) when you click Load transcription model; after that your browser caches it and later visits skip the wait. A recording longer than 30 seconds is cut into 30-second windows that are transcribed one after another: the progress bar fills and the status line counts "window X of N" as each window finishes, and that window's words are added to the output box while the rest keep running. Files up to about 15 minutes are supported - trim longer recordings first.
Use the dictation panel when you can speak the content live, and the file panel when the audio already exists. Clear speech in a quiet room transcribes best; the small on-device model is weaker with heavy accents, overlapping voices, and background noise. Consecutive windows also overlap by two seconds so that no word is cut in half at a seam, which means a word or two can appear twice where one window meets the next. Give the transcript a quick read before you rely on it.
Frequently Asked Questions
What does Speech to Text do?
Convert a SPEECH input into a TEXT output - one step, in the browser.
When should I reach for speech to text?
Typical use: when your downstream platform accepts TEXT but not SPEECH.
What complementary tools work well alongside speech to text?
If you need the reverse direction, look at the related-tools section for the TEXT-to-SPEECH converter.
Is my audio file uploaded when I transcribe it?
No. The file-transcription panel decodes and transcribes the recording in your browser with an on-device Whisper model. The audio never leaves your device; the only download is the model itself, which comes from a public content network and has no access to your file.
Why does the file-transcription panel download about 40 MB first?
That download is the Whisper AI model - the neural network that does the transcription. It is fetched once, only when you click Load transcription model, and your browser caches it afterwards, so repeat visits skip the download.
Why does the status count windows, and why do a few words repeat?
A recording longer than 30 seconds is cut into 30-second windows and each window is transcribed on its own, which is what lets the progress bar and the "window X of N" status line advance as the work goes and add each window's words to the output box. Consecutive windows overlap by two seconds so that no word is sliced in half at a seam, and that overlap is not removed afterwards - so the last word or two of one window can appear again at the start of the next. Delete the repeat when you proofread.
How accurate is the file transcription?
The tool runs whisper-tiny, a small model chosen so it loads fast and works on ordinary devices. It handles clear speech in a quiet room well, but it is weaker with heavy accents, overlapping speakers, and background noise. Proofread the transcript before you use it.