Transcribe Audio to Text: Private speech-to-text with Whisper, right in your browser. Input: Audio, video, MP3, M4A, WAV, OGG, FLAC, AAC, MP4, WEBM, MOV, MKV files and values you enter. Output: Results shown on the page, a downloaded file and text to copy. Privacy: Runs in your browser: inputs and files stay on your device and are not uploaded. On first use it downloads an open AI model to your browser; processing still happens locally.
How to transcribe audio to text
- Drop an audio or video file onto the box: a voice memo, interview, lecture, meeting recording, podcast or video.
- Choose the accuracy. Balanced suits most recordings and is more accurate; Fast is quicker on older phones. The first time, the model downloads once (the size is shown before you start).
- Pick the spoken language or leave it on automatic detection. To get English text from another language, choose Translate to English.
- Press Transcribe. Text appears section by section; click any line to hear that part.
- Download the transcript as plain text or a Word document (.docx), split into paragraphs wherever the speaker pauses (you choose how long a pause counts), or as SRT or VTT subtitles for YouTube, Premiere, DaVinci Resolve or CapCut.
Private transcription with Whisper
Whisper is the open speech recognition model released by OpenAI, trained on hundreds of thousands of hours of speech in nearly 100 languages. This page runs it inside your browser with WebAssembly, so the recording is not uploaded. That makes it suitable for confidential interviews, medical and legal notes, research data, therapy sessions and internal meetings, with no per-minute fees, no account and no upload limits.
Speed depends on your device: a modern laptop transcribes faster than real time with the Fast and Balanced models, while phones are slower. Clear audio with one person speaking at a time gives the best results. Whisper adds punctuation and capitals automatically, but it does not label different speakers, and it can mishear names, jargon and numbers, so proofread important transcripts. Keep the tab open while it works.
Example: accuracy and speed
We transcribed 40 short English recordings and three 4-minute readings from LibriSpeech, a standard public speech test set, with the default Balanced (Whisper base) model in Chrome on a laptop. The word error rate was 4.2% on the short recordings and 3.0% on the long ones: about 3 to 4 words in 100 were wrong. The same base model in whisper.cpp, a popular desktop program, scored 4.4% and 3.9% with its default settings.
A 4-minute recording took about 30 seconds, 7.6 times faster than real time. whisper.cpp is faster on the same laptop: 17 times real time on the processor and 79 times on the graphics chip. The first time, the page downloads the 82 MB model; after that it is kept by your browser.
How it compares with a free transcription service
| This tool | Otter.ai (free Basic plan) | |
|---|---|---|
| Minutes per month | No limit | 300 |
| Length per recording | No fixed limit | 30 minutes |
| Uploaded files | No limit (nothing is uploaded) | 3 imports, lifetime |
Free limits as published by Otter.ai (checked 22 October 2026): Otter.ai pricing. They change often; the paid plans lift them.
Is my audio uploaded?
No. Only the Whisper model is downloaded (once, then cached). Your file is decoded and transcribed on your device.
Which languages are supported?
Whisper understands about 100 languages. English, Spanish, French, German, Italian, Portuguese, Dutch and Japanese work best; choose Balanced for other languages.
Can it make subtitles?
Yes. Download SRT or VTT files with timings split into readable two-line captions, then fine-tune them in a subtitle editor.
How long can the recording be?
There is no fixed limit; long recordings are cut at pauses into pieces of up to 30 seconds. Very long files need more time and memory.
How accurate is free audio-to-text transcription here?
On a standard English test set, the default Balanced model got 3 to 4 words in 100 wrong, slightly better than the same model in the desktop program whisper.cpp. Clear audio with one speaker at a time does best; always proofread names, jargon and numbers.