Free · Private · Local AI

AMR to text converter

Transcribe AMR recordings to text in your browser using a local Whisper model — no upload, no signup, no per-minute fee. Drop one or more .amr files, pick tiny, base or small, and get a timestamped transcript you can copy or export as TXT, SRT, VTT, JSON or DOCX. Language is detected automatically and Chinese output is converted to Simplified. Everything runs on your device.

Your files are never saved. Everything happens on your computer. Nothing is uploaded and nothing is stored on our servers — your audio files stay on your device, and every conversion or transcription runs locally in your browser.
Loading transcription engine…

What people use AMR transcription for

Meeting notes

Turn old phone recordings into searchable, quotable text.

Subtitles

Export SRT or VTT with timestamps for video editors and players.

Batch archive

Queue a folder of recordings — they transcribe one after another, no per-file limit.

Private material

Interview, legal or medical recordings never leave your computer.

How transcription works here

Your AMR file is decoded to 16 kHz mono audio with FFmpeg compiled to WebAssembly. Whisper, running in a Web Worker on your device, then converts the speech to text with word-level timing. Nothing is sent to a transcription API — the only network traffic the tool generates is the one-time download of the model weights and inference engine.

Frequently asked questions

Which Whisper model should I use?

Tiny (~39 MB) runs about 5× faster than real time and suits clear, short voice memos but mishears more words. Base (~73 MB, ~2.5× real time) is the recommended default for interviews and lectures. Small (~238 MB) is the most accurate but runs slower than real time — best for short, important recordings on a modern computer. All three are bundled with the site and cached by your browser after first use.

Can I transcribe several files at once?

Yes. Select or drop multiple .amr files and they are queued and transcribed one after another. The AI model loads only once for the whole batch, and each finished transcript gets its own export buttons. You can also download every TXT and JSON together as a ZIP.

Does transcription work offline?

After the first load, the model files are cached by your browser, so subsequent conversions work without a network connection. The site ships with the model and inference engine, so there is no dependency on an external model host.

Which languages are supported?

Auto-detect identifies the spoken language from the first 30 seconds — including Chinese, Japanese, English and 90+ others — and Chinese results are converted to Simplified characters. You can also force a language from the dropdown if auto-detect struggles with a noisy recording.

What about WebGPU acceleration?

When your browser exposes a compatible GPU, transcription uses WebGPU and runs several times faster. If WebGPU is unavailable or fails for any reason, it automatically falls back to CPU. You can also force CPU-only from the Acceleration dropdown. On Android, CPU is used by default for stability.

Why does it handle long recordings well?

Before recognition, the AMR file is decoded to 16 kHz mono PCM — one channel at a low sample rate. That pre-compression keeps memory use modest, so more minutes of audio fit in the browser than with uncompressed stereo. Very large files (over 100 MB on desktop) can still be slow.

Can it identify different speakers?

No. Whisper produces a single transcript. The text is grouped into paragraphs at pauses, not by speaker. For multi-speaker recordings you will need a diarisation tool in addition.

Is my recording uploaded anywhere?

No. The AMR file is decoded with ffmpeg.wasm and recognised by Whisper running in a Web Worker on your device. The bundled model and engine are loaded from this site; your audio never leaves the browser. Ads, if you have allowed them, are a separate matter and have no access to your files.