AI Ad Detection

Ad detection (sometimes called "ad removal") uses a language model to find host-read ads and sponsor segments in an episode's transcript, then lets the player automatically skip past them during playback. Nothing is cut from the audio file — detected ads are stored as skip ranges, so it's non-destructive and reversible.

Requires the AI container and a language model

Ad detection needs the AI container plus an ad-detection LLM configured in Settings → AI. Unlike transcription (which has a sensible default), ad detection has no default model — you must pick one first. See Setup → Pull a model.

How it works

transcript  ──►  LLM labels ad spans  ──►  mapped to time ranges  ──►  player skips them

1. Ad detection reads the episode's transcript (so the episode must be transcribed first — if it isn't, Pinepods transcribes it automatically as part of the job).
2. The transcript is sent to the configured LLM, which labels which parts are ads.
3. Those are mapped back to precise audio time ranges using the transcript's timings and stored as ad skip segments.
4. During playback, clients seek past any active ad segment.

Because detection works from the transcript, results are approximate to within a few seconds (the boundary of a spoken segment) — which is why skipping, rather than cutting, is the right model, and why you can review and correct results (below).

Turning it on

Manually, per episode

On an episode's Transcript tab, click Detect ads with AI. The job runs (transcribing first if needed) and detected ads then appear in the tab for review.

Automatically, per podcast

In a podcast's settings:

1. Enable Auto-transcribe new episodes (ad detection needs a transcript).
2. Enable Auto-detect ads on new episodes — this option unlocks only when auto-transcribe is on.
3. Choose how detected ads behave:
   • Skip detected ads automatically — ads are skipped as soon as they're found.
   • Require confirmation first — ads are found but not skipped until you confirm them on the episode page.

Reviewing detected ads

Open the episode's Transcript tab. Detected ad ranges are highlighted in the transcript, and a Detected ads list shows each range with Confirm / Deny buttons:

• Confirm — keep skipping this ad.
• Deny — a false positive; stop skipping it (you'll hear it).

Each row shows whether it's currently Skipping or Kept.

Reviews are per-user

Detection runs once per episode and the ad ranges are shared across everyone subscribed to that feed. But whether an ad is skipped — including your Confirm/Deny choices and the auto-skip-vs-confirm setting — is per user. Your decisions never affect anyone else on the server.

Choosing the model

Set this in Settings → AI → Ad-detection model:

• Local (bundled) — runs entirely on your server. Fully private, but slow on CPU and needs a 2–4 GB model file and matching RAM. Good default choice for pull: Qwen/Qwen2.5-3B-Instruct-GGUF → qwen2.5-3b-instruct-q4_k_m.gguf.
• Remote (OpenAI-compatible) — point at an external endpoint such as Ollama or an OpenAI-style cloud API. Usually much faster.
• Remote (Anthropic-compatible) — point at an Anthropic Messages API endpoint (Anthropic itself, or providers that expose that format).

Remote endpoints send transcript text out

With either remote option, the episode's transcript text is sent to that service. Local keeps everything on your server.

Example: z.ai (GLM models)

z.ai sells GLM access in two forms, and they use different endpoints:

Your z.ai plan	Backend	Endpoint URL	Model
GLM Coding Plan (subscription)	Remote (Anthropic-compatible)	https://api.z.ai/api/anthropic	glm-4.6
General API (pay-as-you-go balance)	Remote (OpenAI-compatible)	https://api.z.ai/api/paas/v4	glm-4.5

If you have the Coding Plan, you must use the Anthropic-compatible backend — the OpenAI-style paas/v4 API returns "insufficient balance" because the plan's quota lives on the Anthropic endpoint. Put your z.ai API key in the API key field either way.

Ad detection just classifies transcript text, so a cheaper/faster model tier is usually plenty — you don't need a provider's largest model.

See Setup for how to pull a local model or configure a remote endpoint.

Client support

Client	Ad skipping
Web	✅ Supported
Mobile (iOS/Android)	⏳ Not yet — planned in a future update

Detection and review happen on the server and in the web app today. Mobile skipping is on the roadmap.

Accuracy expectations

LLM ad detection is good but not perfect — host-read sponsorships blend into normal content, and boundaries land on spoken-segment edges (a few seconds of precision). That's exactly why Pinepods gives you per-user Confirm/Deny and a "require confirmation first" mode: treat auto-skip as a helpful default you can correct, not a guarantee.

Troubleshooting

• "Ad removal: Not configured". You haven't chosen an ad-detection model yet. Pull a local model or configure a remote endpoint in Settings → AI, then Save.
• The Detect ads button is missing. The AI container isn't connected (check PINEPODS_AI_URL).
• Auto-detect can't be enabled. Turn on Auto-transcribe for that podcast first.
• It's very slow / high memory. Local CPU inference on a multi-GB model is demanding. Prefer a remote (e.g. GPU-backed Ollama) endpoint for speed.