# Gladia

Gladia pre-recorded transcription jobs and their results, with a Speech tab that uploads, submits and polls a clip in one step.

## What you can manage

- **Workspace** — the API key as a single navigable resource, carrying an activity summary derived from recent jobs and the Speech playground. Read-only.
- **Transcriptions** — pre-recorded transcription jobs with status, file name, audio duration, **billed** time, processing time, detected languages and channel count. The full transcript is fetched inline (delete).

## Credentials

One field. Gladia dashboard at [app.gladia.io](https://app.gladia.io/) → **Account → API Keys**.

The key is sent as the `x-gladia-key` header — Gladia does not use Bearer auth. A single key covers uploads, pre-recorded transcription and the job history; there is no separate admin or usage key.

![Gladia Add-account form with the single API key field](https://agent-assets.infrawrench.com/docs-screenshots/plugins/gladia/add-account.png)

## The Speech tab

Open the workspace for a **Speech** tab. It is transcription only. See [Speech testing](../features/speech-testing.md) for the panel in general.

Gladia is asynchronous, so one press uploads the clip, creates a job, and polls it for up to two minutes. Diarization is on, so speaker labels appear in the word table. Two models are offered: **Solaria-1** (the default, 100+ languages) and **Solaria-3** (latest generation, pre-recorded only).

![Gladia Speech tab on the workspace, showing the Solaria model picker and a diarised transcript with speaker labels](https://agent-assets.infrawrench.com/docs-screenshots/plugins/gladia/speech-tab-solaria.png)

## Tips & limits

- **There is no usage or quota endpoint.** The workspace's activity panel is deliberately not called "Usage": it sums the most recent jobs returned by `/v2/pre-recorded`, which is a **lower bound** on real usage, not a billing figure. The panel says so in as many words, and Infrawrench charts no Gladia spend.
- **Billed time and audio duration are different numbers**, and both are shown on every job. Billed time is the one that costs money.
- **The Speech tab caps clips at 25 MB**, well below what Gladia itself accepts (135 minutes and 1,000 MB). The panel sends audio base64-encoded inside a JSON request, and base64 inflates by a third — 25 MB is what survives that round trip. Send anything larger through Gladia's own API directly.
- **The Speech tab gives up after two minutes.** Long recordings are better submitted from Gladia's own dashboard and read back from the Transcriptions list here.
- **The list endpoint reports no total.** `/v2/pre-recorded` returns `{first, current, next, items}` with no count, so paging walks `next` until it runs out — there is no "N of M" to show.
- **Deleting a transcription returns `202 Accepted`**, so the row may linger for a moment before the next sync clears it.