Transcribe WAV

WAV is uncompressed, and CD-quality stereo runs 1,411 kbps. The median WAV uploaded here runs 129 kbps — close to 8 kHz mono, which is telephone quality. These are dictation machines, call recordings and voice notes saved as WAV, not music.

That is worth saying because it changes the advice. There is no point converting a voice WAV to MP3 to save bandwidth: 1,797 files, median 7 MB and 4 minutes 16 seconds, upload in seconds as they are. And there is no accuracy penalty either way — the model resamples everything to 16 kHz mono.

These are medians measured on our production cluster, expressed as time per hour of audio.

Against the WAV files uploaded here: the median, just over four minutes, comes back in about six seconds. Nine in ten are under 42 minutes.

No. WAV holds the samples untouched, which is the right thing for an archive or an edit, but the transcriber works at 16 kHz mono and everything above that is discarded before the model sees it. A WAV and a 128 kbps MP3 of the same recording arrive as very nearly the same signal.

The old claim on this page — 98% accuracy for WAV against 95% for compressed formats — was never measured and has been removed. What does decide accuracy is the recording: microphone distance, room, overlapping speakers, and bitrates under about 32 kbps.

It depends on what kind of WAV you have, which is why both limits are worth knowing. At voice bitrates duration binds long before size. At CD-quality stereo, 5 GB is about eight hours — the two limits arrive together. At 96 kHz/24-bit, size binds at roughly two hours.

• The largest WAV processed here was 2.84 GB. Nothing has hit the ceiling yet.
• Converting to FLAC halves the upload and is lossless — the samples decode back identically. Converting to MP3 also works and costs nothing in transcript quality.