How to transcribe AMR voice recordings to text
Voice notes are quick to record and painful to search. Here is how to turn AMR recordings into text — including the privacy trade-offs most guides skip.
Why transcribe at all
A spoken sentence takes seconds to record and a long time to skim. Text is searchable, quotable and easy to paste into notes, reports or subtitles. Transcription is especially useful for old recordings you need to audit — interviews, meetings, voicemails — where listening through each file is not realistic.
AMR is a slightly awkward input for this. It is 8 kHz mono speech, which is actually fine for recognition: speech models are trained on exactly this kind of signal. The low sample rate is not a problem; background noise and overlapping speakers are.
Cloud APIs versus local transcription
Most transcription services work by uploading your audio to a server. That is convenient and often accurate, but it means a private conversation leaves your device and a per-minute charge accumulates. For interviews, legal material or medical notes, that trade-off may not be acceptable.
The alternative is running a speech model locally. Modern browsers can do this with models such as Whisper compiled to WebAssembly or WebGPU. The AMR to text tool on this site works this way: the model downloads once, then recognition happens entirely in your browser. Nothing is uploaded and there is no per-minute fee.
The cost is speed. A local model runs at roughly real-time to several times real-time depending on your machine and the model size, so a one-hour recording takes minutes rather than seconds. For most personal archives that is a fair trade.
Choosing a model size
- Tiny (~39 MB): fastest, useful for short clear voice notes, but mishears more words.
- Base (~73 MB): the sensible default for interviews and lectures; a good balance of speed and accuracy.
- Small (~238 MB): noticeably more accurate on difficult audio, but slower than real time on many computers.
Start with Base. Move to Small only if the recording matters enough to justify the wait, or drop to Tiny if you are processing a large batch of clean, short memos.
Getting better accuracy
Transcription quality is mostly decided by the recording, not the model. A few things genuinely help:
- Reduce steady noise first. Air conditioning, fan hum and tape hiss confuse recognition. Running the file through the noise reducer before transcribing can cut errors.
- Even out the volume. Quiet passages are where models hallucinate words. Normalising loudness helps.
- Pick the right language. Auto-detection works well from the first 30 seconds, but forcing the language in a dropdown is more reliable on noisy or very short clips.
- Accept the limits. Whisper produces one transcript, not a speaker-separated conversation. Overlapping talk will be merged into a single stream.
Working across languages
Whisper supports around 100 languages, and modern implementations can identify the spoken language automatically. This matters for AMR archives from phones used in different countries — a folder of voice notes may mix languages.
One detail worth knowing: the model sometimes emits Traditional Chinese characters even when the speaker uses Mandarin. A local conversion step to Simplified Chinese, as used by the transcription tool here, keeps the output consistent without another service.
Export formats and what they are for
- TXT — plain reading copy, easiest to search and quote.
- SRT / VTT — subtitles with timestamps for video editors and players.
- JSON — structured text with timings and detected language, for further processing.
- DOCX — a Word document when you need to hand the transcript to someone who does not use plain text.
Export more than one format the first time. What looks like a simple note often ends up in a video project or a report a week later.
Handling long recordings and batches
Very long recordings deserve a plan. Local transcription holds audio in memory, so a two-hour file is slower and uses more RAM than several shorter ones. If a recording has natural breaks, splitting it into sections first makes the job faster to retry and the result easier to check against the original audio.
For a whole folder, queue the files and let the model load once for the batch. Accuracy on a long session improves more from a quiet source and the correct forced language than from switching model sizes. If a passage comes out wrong, replay that section rather than re-running everything — targeted correction is quicker than chasing a perfect transcript in one pass.
A realistic workflow
For a folder of old AMR recordings: open the transcription tool, queue the files, pick Base, and let it run. Clean up the worst noise first if the recordings are poor. Download the TXT and JSON together as a ZIP, keep the originals, and use the results as the searchable index of your archive. No upload, no monthly bill, and the audio stays yours.