What audio formats Pincite Audio accepts

Recorded-call audio does not arrive as a podcast file. It arrives as whatever the recording system wrote — a correctional vendor's export, a carrier or phone system, a handheld — which is usually a narrowband telephony codec chosen decades ago to carry intelligible speech over a constrained line — not to preserve a recording for later analysis.

That shapes what a transcriber can do with it, and it is worth knowing before the file reaches one.

What has been driven through the media path

Six codec families have been run through the production media path — real ffmpeg and ffprobe, the shipped command lines, with only cloud storage swapped for a local directory: μ-law and A-law PCM, Microsoft ADPCM, GSM 6.10, AMR narrowband, and dual-channel G.729. That run covers reading the file, planning and applying normalization, and splitting it for transcription. It does not cover transcription itself.

Two things about that exercise are worth stating, because they are the difference between evidence and a claim:

• Every fixture was synthesized for the test — thirty seconds each, not a real recording. The G.729 fixture in particular carries a deterministic byte pattern rather than speech, so that case proves the container handling and the decode invocation and nothing about how a real G.729 call transcribes.
• All six were exercised at 8 kHz, and two of the six as two-channel files. That is one sample rate, not a range.

So this is an evidence inventory, not a compatibility contract. It says these formats have been proven through the media path. It does not say they are the only ones that work, and it does not say anything about any particular vendor's export.

There is no extension allowlist on upload

Nothing blocks a direct upload on the basis of its name. That surface does not filter by extension, suffix, or declared content type — the probe reads the actual file and decides whether it can be handled.

The practical consequence is the useful one: a file with an unfamiliar or missing extension is not rejected for that reason, and a file with a familiar extension that is actually damaged does not slip through on the strength of its name.

Pulling from Clio works differently. That browser only offers documents whose extension or content type marks them as audio or video, so a mislabelled file is refused there before any bytes move.

What happens to a file it cannot read

It fails at the probe, with an operator error code, before any transcription is attempted.

That is a design choice rather than a limitation. The alternative — decoding what can be decoded and transcribing that — produces a document that looks finished and is silently missing whatever failed to decode. A file that fails loudly is a file you know to go back for. A partial transcript that nobody flagged gets read as complete.

Upload the original

The instinct to convert a strange file to something familiar before uploading is the wrong one here. Every lossy re-encode discards information, and an already-narrowband recording has little to spare. Converting a compressed 8 kHz file to MP3 does not recover anything; it removes a little more.

Normalization happens once, inside the pipeline, from the file you provided — and telephony-band audio keeps its source sample rate rather than being resampled upward, because upsampling invents no detail. Sending the original is what gives that step the most to work with.