↑↓ to move · Enter to open · Esc to close

Audio Transcription

Turn speech into editable text on a compatible WebGPU device. The multilingual Whisper Tiny model runs in your browser, with a maximum accepted recording length of 15 minutes.

Audio and transcripts stay on your device. Processing runs in your browser.
Built and reviewed by · Updated

Choose a file

MP3, WAV, M4A, OGG, FLAC, MP4. Up to 100 MiB per file (50 MiB on touch devices), 900 seconds. Decoded audio memory limits also apply.

    Settings

    Checking WebGPU support…

    The multilingual Whisper Tiny model downloads on first use from Hugging Face and is cached when storage permits. Audio stays here. Accuracy varies; check names and punctuation before exporting.

    Result

    Your result will appear here

    Add a file, pick your settings, then select Transcribe audio.

    No account or daily usage quota. File and device limits protect your browser from running out of memory.

    Processing audio

    Checking your settings…

    How to use Audio Transcription

    1. Check that the page reports an available WebGPU adapter, then choose a supported recording.
    2. Select the spoken language or leave automatic detection on. Persian and English are available language choices.
    3. Press Transcribe audio. The speech model downloads on first use and may be cached by your browser.
    4. Review the transcript and timed segments. Correct names, punctuation and any recognition errors.
    5. Copy the text or download TXT. Export SRT or VTT when reliable timed segments are available.

    Local speech recognition has device requirements

    The audio and transcript stay on your device. JavaScript and the inference runtime load from this site; multilingual Whisper Tiny model files download from Hugging Face when needed. Model downloads use bandwidth and can take time before the first transcript starts.

    The browser needs HTTPS or localhost, working WebGPU and enough memory for decoding and inference. A detected GPU adapter is a capability check, not a promise that a long recording will finish. Phones can stop work when the screen locks or the browser suspends the tab.

    Review text and subtitles separately

    The editable transcript is used for copy and TXT export. Timed segment fields are used for SRT and VTT, so changing the main transcript does not silently alter subtitle timing. Edit segment text when you need to correct a subtitle.

    Audio is resampled to 16 kHz mono and processed in sequential overlapping windows. Overlap is reconciled using segment timestamps to reduce repeated text at boundaries. Music, long silence, background speech and poor recordings can still cause omissions, invented words or timing errors.

    A 15-minute maximum is not a speed guarantee

    The accepted duration can be lower when the site’s file or memory settings require it. Fifteen minutes is the product ceiling, not a guarantee of real-time processing or compatibility with every device. Try a short sample first and keep the tab open.

    You can cancel inference or unload the model. Model files may remain in browser storage for later use. If a source needs trimming before transcription, MP3 Cutter can prepare an MP3 excerpt.

    Before you download

    • WebGPU is required. Unsupported devices receive a message; the tool does not upload audio for a server fallback.
    • Whisper Tiny can make recognition and timestamp errors. Human review is needed, especially for names, specialized terms and noisy speech.
    • SRT and VTT export require usable timed segments. The tool does not invent subtitle times for untimed text.
    • Long recordings can exceed GPU or browser memory. Cancel or unload the model, then try a shorter file if processing fails.

    Frequently asked questions

    Is my audio sent to an AI service?

    No. Inference runs locally. The browser downloads model files from Hugging Face, but it does not send your selected media or transcript there.

    Can it transcribe Persian?

    The selected model is multilingual, and Persian is a language option. Accuracy varies by recording and device; review the output rather than relying on it as an exact transcript.

    Can I transcribe a full 15-minute recording on any phone?

    No. Fifteen minutes is the maximum accepted duration. Actual support depends on WebGPU, available memory, file settings and the browser staying active.

    Why do subtitle exports differ from my edited transcript?

    TXT and copy use the main text field. SRT and VTT use timed segment fields. Edit those segment fields to correct subtitle text while preserving the detected times.

    Often used together with Audio Transcription.

    • MP3 Cutter

      Trim MP3 audio with waveform positioning and exact time fields.

    • Audio Converter

      Convert between common audio formats with bitrate and channel controls.

    • MP4 to MP3

      Extract an MP3 soundtrack from a local MP4 video.