AI Transcription

With the AI container running, Pinepods can generate a timecoded transcript for any episode using Whisper speech-to-text. This is separate from built-in transcripts that some podcasts publish in their feed — both show up in the same place.

Requires the AI container

Transcription needs pinepods-ai running and PINEPODS_AI_URL set. Without it, the transcription controls are hidden. See Setup.

Ways to transcribe

Manually, per episode

On an episode's page, open the Transcript tab and click Generate with AI. A job is queued and you'll see live progress (transcribing a long episode takes a while on CPU). When it's done, the transcript appears in the tab.

• You don't need to have downloaded the episode first — the server fetches the audio temporarily, transcribes it, and cleans up.
• You can Regenerate at any time to replace an existing AI transcript.

Automatically, per podcast

In a podcast's settings, turn on Auto-transcribe new episodes. From then on, new episodes for that podcast are transcribed automatically as they arrive.

• This is opt-in per podcast — there is no "transcribe everything" switch.
• It only affects new episodes going forward, not your back catalog.

Reading the transcript

The Transcript tab shows the transcript as timecoded segments. Each line is timestamped, and clicking a timestamp jumps the player to that point — handy for skimming or finding a specific moment.

If an episode has both a built-in (feed-provided) transcript and an AI-generated one, a small source selector lets you switch between them.

How it works (and what's shared)

• Transcripts are content-level: a given episode is transcribed once, and the result is shared across everyone subscribed to that feed on your server. It isn't redone per user.
• The transcript is stored in the Pinepods database, so it's instantly available on later visits without re-running Whisper.
• Both the full text and per-segment timings are stored, which is what powers the timecoded view — and what ad detection uses to map detected ads back to precise time ranges.

Choosing a model

The Whisper model is set in Settings → AI → Models. Trade-offs:

Model	Speed	Accuracy	Notes
tiny / base	Fastest	Lower	base is the default; great for most uses
small / medium	Slower	Higher	Noticeably better on tricky audio
large-v3	Slowest	Best	Heavy; best with a GPU

The first time a model is used it downloads automatically and is cached on the /models volume.

Tips & troubleshooting

• It's slow. That's expected on CPU. A one-hour episode can take many minutes. Use a smaller model, or a GPU-enabled setup, if speed matters.
• The button is missing. The AI container isn't connected — check PINEPODS_AI_URL and that the sidecar is healthy (Settings → AI shows the connection status).
• Nothing transcribes automatically. Auto-transcribe is off by default; enable it per podcast.