This tool turns recorded speech into text without sending the recording to a transcription service. The Whisper model runs inside a browser worker, keeping the page responsive while your computer does the inference. WebGPU is used when available; other browsers fall back to a slower CPU mode. The downloaded model is cached by the browser for later visits.
Audio to Text Converter
Runs entirely in your browser - no upload, no sign-up.
How can I convert audio to text privately?
Add an audio file and this audio to text converter runs the Whisper speech recognition model on your device. It produces an editable transcript with TXT and SRT exports. Your recording is never uploaded, although the browser must download the AI model the first time you use it.
How to use
- 1Choose the recording. Drop an MP3, WAV, M4A, OGG, FLAC, or WebM file onto the tool, or click the upload area to choose one.
- 2Run private transcription. Click Transcribe audio. On the first run, wait while the multilingual Whisper model downloads, then let it process the recording on your device.
- 3Review and export. Listen back to uncertain sections, edit any mistakes, then copy the transcript or download it as TXT or timestamped SRT.
Who it's for
- Interviewers and researchers creating a private first draft before checking names and quotations against the recording.
- Students turning recorded lectures or study notes into editable text without uploading classroom audio.
- Podcasters and video creators producing a rough transcript or timestamped SRT file for captions.
- Teams handling sensitive recordings keeping source audio on the device instead of sending it to a cloud transcription queue.
FAQ
Is my audio uploaded to a server?
No. The browser decodes your selected file and sends the audio samples only to a local worker on the same device. The worker downloads the Whisper model files from Hugging Face, but it does not upload your recording or transcript. CanDoYa does not receive or store the file.
Is the audio to text converter free?
Yes. There is no account, payment, transcription credit, or watermark. Your own device performs the AI inference, so the practical cost is processing time, memory, and the first model download. Your internet provider may still count the model download against a metered data allowance.
What file size and recording length can I transcribe?
The upload control accepts files up to 500 MB. There is no fixed minute quota, but long recordings need more memory and take longer because inference happens on your device. A modern laptop is a better choice than a phone for hour-long meetings or lectures.
Which audio formats are supported?
The tool accepts MP3, WAV, M4A, OGG, FLAC, WebM, and several related formats. Actual decoding support depends on your browser and operating system. If a file is rejected even though its extension is listed, convert it to MP3 or uncompressed WAV and try again.
Which languages can Whisper transcribe?
The selected Whisper tiny model is multilingual and automatically detects many spoken languages. Accuracy varies by language, accent, recording quality, and topic. It does not translate the transcript into another language, identify speakers, or guarantee correct spelling of names and specialized terms.
Why does the first transcription download a model?
Private on-device transcription needs the speech recognition model on your computer. The WebGPU path downloads about 80 MB, while the compatible quantized CPU fallback is about 105 MB. Browsers normally cache those files, so later visits can reuse them unless site data or the cache is cleared.
Does audio transcription work without WebGPU?
Yes. The tool prefers WebGPU because a supported graphics processor can run Whisper much faster. When WebGPU is missing or fails, it retries with WebAssembly on the CPU. That fallback works in more browsers but may be slow for long recordings, especially on phones and older computers.
How accurate is the transcript?
Whisper tiny gives a useful draft from clear speech, but it is not a verbatim guarantee. Background noise, music, crosstalk, weak microphones, uncommon names, and long silent sections reduce accuracy and can cause hallucinated words. Review important quotes, medical details, legal statements, and decisions against the original audio.