gratistools.be
Source code Sync (optional)

Private AI Transcription & Subtitle Studio

Transcribe recordings up to two hours locally. Choose faster Whisper Tiny or more capable Whisper Base, filter rumble or hiss and export text or subtitles.

Local: yes. Your input stays on your device and is not sent to an external service for this tool.

Transcribe with local Whisper

Long MP3 recordings are decoded into short sections using local FFmpeg. A multilingual Whisper model is downloaded on first use. On phones, the more compatible CPU engine is selected automatically; on desktops, WebGPU can be faster. Your audio stays on this device. Keep this tab open.

For noisy Dutch, select Dutch and try Base. Cleanup is mild and optional: it cannot remove overlapping voices or recover inaudible speech. Base downloads more model data directly from Hugging Face.

Detailed processing status and errors

Every percentage refers to an individual model file. The engine is only ready when it says Whisper initialized. A stalled engine is stopped and retried on CPU when possible.

Choose a recording to see diagnostic steps.

Partial transcripts remain downloadable after each section, even if the model fails later. Stop or resume in this tab. Do not close or reload before exporting.

Recording and transcript stay on this device. FFmpeg and the ONNX runtime are served by GratisTools; Whisper Tiny/Base model weights come from Hugging Face. CPU works without WebGPU but can be slow, especially with Base. Check transcript and timestamps.

Choosing accuracy and language

Whisper Tiny is the faster default; Whisper Base uses a larger multilingual model that may recognize difficult speech more accurately, but needs more memory, a larger download and more processing time. For a Dutch or Flemish recording, choose Dutch instead of automatic detection. Other explicit languages are English, French and German.

Optional audio cleanup

The optional light filter attenuates low-frequency rumble and high-frequency hiss in each audio section before recognition, entirely on your device. It is not speech separation or AI noise removal and may not improve every recording. You can leave it disabled and compare results.

Transcribe long MP3 recordings

Process audio up to two hours with a maximum compressed file size of 350 MiB. FFmpeg decodes one short section at a time, so the browser does not hold two hours of raw audio in memory. The inference runtime runs in a separate worker, and the page reports individual download, startup and transcription stages. Processing may still use substantial memory and take considerable time. The default uses CPU/WASM on phones and WebGPU on supported desktops, falling back to CPU if GPU startup or transcription stalls. CPU may be much slower. Files larger than 350 MiB must be compressed or split.

Privacy

The open-source Transformers.js runtime downloads a multilingual Whisper model and performs recognition on the device. FFmpeg is served from this site. Your original audio is not sent to a transcription API. You must keep the tab open for long processing jobs.

Progress, pause and resume

Progress is updated after each short section. Stop after the current section and download partial text or subtitles, or resume remaining sections in the same tab. Reloading the page loses unsaved results. Choose the spoken language to improve consistency over long recordings.

Subtitle accuracy

Review the transcript and individual timecoded segments before downloading SRT/VTT; timestamps and words are approximate. Editing the plain text does not automatically retime captions.

Frequently asked questions

Is this tool free?

Yes. The core tool is free to use in your browser.