Transcribe FLAC

It does not, and this is the single most useful thing to know before uploading a FLAC archive. Speech recognition models work on 16 kHz mono audio: Whisper builds its input from a log-mel spectrogram at that rate, and everything above 8 kHz — half the sample rate — is gone before the first layer of the network.

So a 96 kHz/24-bit master and a 44.1 kHz/16-bit CD rip of the same recording arrive at the model as the same 16 kHz mono signal. The difference you can hear on good headphones is real; the difference the transcriber sees is nil.

Where audio quality does change the transcript

• Very low bitrates. Below roughly 32 kbps, lossy codecs start smearing consonants, and that survives the downsampling. This is why a bad voice note transcribes worse than a good one.
• Repeated re-encoding. Each generation adds artefacts. FLAC has no generations — decode and re-encode as often as you like and the samples are identical.
• The recording itself. Microphone placement, room, overlapping speakers and background noise dominate everything else. No container fixes any of them.

FLAC is the right format to keep a recording in. It is not a way to get a better transcript out of one.

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

FLAC's size does not appear in that table because it does not belong there. The file is decoded once on arrival, and from that point a lossless hour and a 128 kbps hour are the same number of samples. What FLAC does cost is upload time: at roughly 350 MB per hour of CD-quality stereo, a two-hour recording is around 700 MB going up the wire, where the same audio as MP3 would be under 60 MB.

If upload speed is the bottleneck and the archive is safe elsewhere, converting to a decent MP3 before uploading is a reasonable trade — it will not cost you transcript quality. If the FLAC is the only copy, upload it as it is.

FLAC is an archive format, so the files that reach us tend to be masters rather than working copies.

1. Field recorders. Zoom, Tascam and Sound Devices units write WAV natively; FLAC is what people convert to for storage. Both upload here unchanged.
2. Audacity and Reaper. File → Export → Export as FLAC in Audacity; any bit depth works.
3. Broadcast and institutional archives. FLAC's patent-free licence is why libraries and radio archives standardised on it — an archive you can still open in thirty years without asking anyone's permission.
4. CD rips. Interviews and lectures pressed to disc, ripped losslessly rather than re-compressed.

Whichever it is, drop the file into the box at the top of this page. The first five minutes need no account.

FLAC is the format where duration binds before size — the opposite of the video containers. CD-quality stereo FLAC runs about 350 MB per hour, so the 5 GB ceiling holds roughly fourteen hours, while the duration limit stops at eight.

• Split by session, not by clock. A day of interviews is several recordings, not one eight-hour file. Cutting between them costs nothing; cutting mid-sentence costs the sentence.
• High-resolution masters are the exception. At 96 kHz/24-bit stereo a FLAC runs closer to 1.2 GB per hour, and there the 5 GB limit arrives around four hours instead. Downsampling to 44.1 kHz before uploading is lossless as far as the transcript is concerned — the model resamples to 16 kHz regardless.
• Keep the original. Whatever you upload, keep the untouched master. A transcript can be re-run; a discarded original cannot be recovered.