Pincite Audio

Transcripts your audio evidence can actually be cited from.

Recorded calls, interviews, and recorded statements turned into a transcript that discloses what it couldn't hear, never guesses a name, and carries a SHA-256 chain-of-custody record.

Create an account

The problem

What consumer transcription leaves out

Consumer transcription hands back a wall of text and SPEAKER_00. It is cheap and it is fast, and it still leaves you listening to the audio — because you cannot tell which parts it skipped, which names it guessed, or which sentence you can safely quote in a filing.

The same stretch of one recording, both illustrations, speech content not shown. On the left, the twenty-seven seconds that were never transcribed are not marked — there is nothing on the page to tell you they are missing, and SPEAKER_00 reads the same whether the machine knew who was speaking or guessed. On the right, a recording from a single microphone attributes every line to one unidentified speaker rather than splitting it among guesses. A gap is a fact about one channel: where a recording's channels were recorded separately, the marker names the one it belongs to — [NO TRANSCRIPTION RETURNED ON CHANNEL 2 00:04:12–00:04:39].

What's different

What's different

Nothing silently missing

Every second the system returned no transcript for is marked on the page, in every export — except stretches it measured as silence, which are recorded but not rendered. A coverage measurement runs against the source file, and untranscribed spans render as explicit gaps rather than being dropped. Where a stretch came back nearly empty instead of empty, there is nothing on the page to mark: those seconds are named separately instead, on the page and on the certificate, and the transcript is still delivered.

No name the record can't support

A speaker label always carries where it came from. Labels derived from a call's channel describe a role, never a proper name, and a speaker the system could not verify has no name at all — enforced in the database, not by convention.

A transcript that carries its own digest

Each transcript carries a SHA-256 chain-of-custody record over three things: the source file exactly as your firm delivered it, the audio submitted for transcription part by part, and the produced text. A recording ingested before source hashing says so on the certificate's face rather than leaving the field blank.

Private by architecture

Every matter's transcript is encrypted under its own key. No CDN or third-party proxy sits in the path where transcript text is decrypted.

An illustration of the transcript format — not a real transcript. Speech content is not shown. The gap marker is the literal string the system renders, and the speaker designations are the literal shapes it produces. A gap is a fact about one channel: this recording's channels were recorded separately, so the marker names the one it belongs to; a recording without channel designations renders the same marker with no channel named. See what the output looks like.
An illustration of the matter listing — not a real matter. Recording names are redaction bars and no coverage percentage is shown. The verdict lines and the word counts are the literal strings the product renders, with illustrative numbers: every recording leads with what is known about its completeness, and a recording the system could not measure says so in words rather than showing a figure.

Create an account →

How it works

How it works

  1. Send us the audio.

    Recorded-call exports and interview recordings — in whatever format it arrived in: a county jail vendor's export, a phone system, a handheld recorder.

  2. We transcribe it and mark what we couldn't.

    Channels are separated, speakers are labeled with their provenance, and gaps are disclosed on the page.

  3. You get something you can quote.

    A transcript with stable references, and a chain-of-custody record for the audio that was transcribed.

In Claude

Then ask it questions, in Claude.

Claude cannot open an audio file — no Anthropic surface accepts one. Pincite Audio connects to claude.ai as a custom connector, so the transcript is what reaches Claude. The audio does not.

Read-only, by scope

The connector asks for one permission — read transcripts — and every tool behind it is a read. There is no tool that writes to a matter, edits a transcript, or changes a speaker label.

Every line comes back cited

A passage is returned with a citation on each line, generated by the server rather than written by the model, and the quoted text is byte-identical to what is stored. You are not retyping a quote out of a chat window.

"Nothing found" is answered, not shrugged

A search that finds nothing does not come back as a bare no. It comes back flagged, with how many recordings in that matter carry gaps, and a second tool reports where those gaps are and between which times.

Your firm's matters, and no others

A connector session reaches the matters its firm can access. The same wall the application enforces is the wall the connector sits behind.

One limit, stated plainly: once a passage is in a Claude conversation it is in your Anthropic account, not ours. We record which tool was called and how much came back — never the transcript text, and never the conversation.

Who it's for

Built for criminal defense and small litigation firms.

Pincite Audio is for 1–5 attorney firms — criminal defense, employment, plaintiff-side, personal injury, family, and civil rights — in jurisdictions where the record arrives as media rather than paper.

Speakers are separated by channel when the recording keeps them on separate channels. On a single microphone the recogniser groups the lines by voice instead — a machine guess, not an identity: the product does not establish how many people spoke, or who any track is, and you name the tracks yourself.

It is not the cheapest way to transcribe an hour of audio, and it is not trying to be. If you need bulk transcription at the lowest price per hour, use Rev. Pincite Audio is for the recordings where it matters that the transcript is complete, attributable, and citable.

Create an account →

Architecture

What we do with your audio.

Not ready to create an account?

Leave an email and we will contact you about getting started. We will use it for nothing else.

Submitting shows a short confirmation page — there is no automated email; we read the list ourselves and write back. How the address is handled is described in the privacy notice.