The format of phone calls and older recorders — the hardest material we accept, and worth knowing why.
You convert an AMR file to text by uploading it without conversion. Honestly: it is the hardest material we accept, because AMR-NB writes an 8 kHz band at a few kilobits per second — enough for a phone call and no more. Recognition will work, but expect more mistakes than from a dictaphone or a phone held to the mouth.
Drop a file here or choose one from your device. We will show you the beginning of the transcript first.
skryba.ai transcribes on Scribe v2, which gets 2.2% of words wrong in Artificial Analysis' independent index — the lowest of every model measured. Gemini 3.5 Transcribe gets 2.6% wrong, Whisper large-v3 4.1%. It is the same model that writes the transcript preview you see before paying.
| Model | WER |
|---|---|
| skryba.ai (Scribe v2)the model we transcribe on | 2.2% |
| MAI-Transcribe-1.5Microsoft Azure | 2.4% |
| Gemini 3.5 TranscribeGoogle | 2.6% |
| Universal-3 ProAssemblyAI | 3.1% |
| GPT TranscribeOpenAI | 3.3% |
| Whisper large-v3OpenAI | 4.1% |
| Nova-3Deepgram | 5.2% |
Source: Artificial Analysis, AA-WER v2 index, non-streaming mode — read 27 August 2026. Seven of the 40-plus models measured: the top of the table and the names you already know. The index is measured mostly on English audio; for other languages, Polish included, no comparable public ranking exists. That is why we show you the start of every transcript before we ask for anything.
AMR was built to fit human speech into a GSM channel, not to record it faithfully. In its narrowband form it samples at 8 kHz with a bitrate between 4.75 and 12.2 kbps — for comparison, an ordinary dictaphone MP3 has several times the bitrate at twice the bandwidth.
The practical effect is always the same: the high frequencies that separate sibilants disappear. In Polish “s”, “sz” and “ś” suffer most, and after them the inflectional endings — precisely what carries a sentence's grammar. Surnames and proper nouns come out worse than the rest, because context cannot help guess them.
We say this plainly because the preview is free, and it is better that you judge the quality on your own recording before starting a trial rather than after. If the result does not convince you, you pay nothing.
For next time: most phones today can record with the system voice recorder, which writes full-bandwidth M4A. If a conversation is going to be transcribed later, that is the difference between hard material and easy material — not a question of better transcription software.
A call recording is usually somebody else's words too. It is worth being clear on what basis it was made and who you share the finished transcript with — we store recordings privately, and your account settings can delete the audio file automatically after 1, 7 or 30 days.
It is, just with lower accuracy than an ordinary recording. The preview is free and needs no account, so the cheapest way to find out is on your own file — a glance at the first few sentences is enough to judge whether the result is usable.
No. Converting enlarges the file without adding information that is not in it — an 8 kHz band stays an 8 kHz band. Upload the original.
We take that too. 3GP is a container from older phones, and the audio inside it is often AMR — sometimes with a picture, sometimes alone. Upload it without converting.
