Deepgram

Deepgram (/media/deepgram) is a bespoke, non-generation studio. It transcribes speech in an audio or video file from your Media Library into text, word timings, and caption segments. It does not produce an image/video/audio artifact, so it does not follow the standard Studio Kit generation flow.

Configure your Deepgram API key at Settings → Media. There is no environment-variable fallback; if the provider is not configured, the studio shows the standard "isn't configured" empty state.

Pick a source

1. Click Pick audio / video and choose a file already in /files.
2. The source must be a library file (the backend needs a fileId). External URLs are not accepted.
3. Accepted containers/extensions include mp3, m4a, aac, wav, ogg, flac, webm, mp4, mov, and m4v.

Transcription options

• Model — nova-2 (default), nova-3, or whisper.
• Language — optional BCP-47 code (for example en, es, de). Leave blank to let Deepgram auto-detect.

How transcription works

When you click Transcribe, the backend reads the source file's bytes directly from storage (local or cloud adapter). The audio never leaves the deployment as an outbound HTTP URL, so there is no SSRF surface for this step.

The request is sent to Deepgram's /v1/listen endpoint. Smart formatting and punctuation are enabled by default; a language code is passed only when you supply one. Sources larger than 250 MB are rejected before transcription starts.

The response returned to the panel is:

{
  "text": "the full transcript",
  "words": [{ "word": "the", "start": 0.12, "end": 0.18 }, ...],
  "segments": [{ "start": 0.12, "end": 2.45, "text": "the full transcript" }, ...]
}

segments are phrase-chunked on sentence-ending punctuation or after 12 words, matching the grouping used by the Designer timeline's auto-caption feature.

After transcribing

The result panel shows an editable transcript and a timecoded segment list. The action bar offers:

• Copy — copy the transcript text.
• Download .srt / Download .vtt / Download .txt — client-side exports; nothing is written to /files.
• Save to Files — persists the transcript as a completed stt AIMediaJob via POST /media/deepgram/save-transcript. Because it is created already complete, it never enters the async poll path. In the render queue it appears as a text card with Copy and To composer buttons only — no preview, no Edit-in-Designer/Post handoff.
• Send to composer — opens the composer pre-filled with the transcript text.
• Edit in Designer — for a video source, stashes { url, fileId, width, height, words } and opens /media/designer?captions=1. The Designer builds a video project from the clip and burns in a caption track from the word timings, with no re-transcription. For an audio source, the clip is opened directly on the timeline's audio track instead.

Related docs

• Media Studios — where this studio fits relative to the standard generation flow.
• Designer — manual canvas and video timeline editor.
• Media Library — /files upload and folder management.
• Settings — configuring the Deepgram media provider.

---

Verified against v1.0.0