Audio Transcription
Turn recorded speech into editable text on a compatible WebGPU device. Recordings must be 15 minutes or shorter. The speech model downloads before use, and the generated transcript needs review.
Choose a file
MP3, WAV, M4A, OGG, FLAC, MP4, WEBM. Maximum 15 minutes per file. 100 MiB on desktop, 50 MiB on touch devices.
Settings
A working WebGPU device is required. The first run downloads a multilingual speech model. Recordings up to 15 minutes are accepted on supported devices; processing speed and accuracy vary. TXT is always available after success; SRT/VTT require usable timestamps.
Result
Your result will appear here.
Files and generated text stay on this device. Processing code is downloaded from ICANDL; speech model files are downloaded from Hugging Face. No media upload or server fallback is used.
This tool was developed by Utilza Team.
How to use Audio Transcription
- Choose a supported recording up to 15 minutes on a compatible WebGPU device.
- Select the available language option and let the speech model download.
- Transcribe the recording, review the text and save your corrected result.
Check device support before transcription
Local speech recognition requires a working WebGPU connection and enough device resources for the selected model and recording. The first run needs an internet connection to download model files. A 15-minute maximum is an input limit, not a promise that every device can complete a recording of that length.
Review the transcript against the recording
Recognition can confuse names, accents and technical terms. Background noise, overlapping speakers and silence can also produce omissions or invented words. Listen to the source and correct the text before using it. Your recording and transcript are processed locally; the model assets are separate downloads.
Frequently asked questions
Does transcription work on every phone or browser?
No. It requires compatible WebGPU support and sufficient device resources. A browser that cannot run the model receives an unsupported-device message instead of a server transcription fallback.
Why does a model download before transcription?
The browser needs speech recognition model files to run local inference. The first download requires internet access and may take time; browser storage availability affects whether cached files can be reused.
How accurate is the generated text?
Accuracy is not guaranteed. Language, recording quality, speakers and vocabulary affect results. Review and correct the transcript against the audio.
Can I transcribe more than 15 minutes?
Each selected recording must be 15 minutes or shorter. Shorter clips can still exceed device resources, and an accepted extension must contain a supported audio codec.