How a sermon recording becomes a study guide. The AI pipeline, the human review, and how to flag a correction.
How a sermon becomes a study guide
Every published sermon on Sunday is processed through a four-phase pipeline. The pastor records the sermon as usual; everything below happens after upload, in roughly four to seven minutes per hour of audio.
01
Transcription
The audio is transcribed into words with timestamps, sentence by sentence. Whisper Large handles English; non-English audio falls back to a slower multilingual variant. Filler words and stutters are preserved at this stage — we do not yet edit.
OpenAI Whisper-Large-v3
02
Reading the sermon
A summary, key themes, and clip candidates are extracted from the transcript. The model is told to read the sermon as a careful study tool, not summarize it as if for a press release. Identified clips are checked against sentence boundaries so they start and end at a natural breath.
OpenAI GPT-4o
03
Scripture retrieval
Bible references the pastor cited are matched against canonical text in our scripture index. Each reference is tagged as either explicit (the pastor preached on this passage) or related (mentioned in passing or thematically connected). Only explicit references are used on cross-church scripture pages.
OpenAI embeddings + Pinecone vector index
04
Discussion questions and review
Discussion questions are generated in three categories: observation, interpretation, and application. A second model independently reviews every question and every scripture reference, rewriting or removing items that fail a quality bar.
OpenAI GPT-4o · Anthropic Claude Sonnet 4.6 (judge)
The models we use
We use third-party language models from OpenAI and Anthropic. We do not train our own models. The audio you upload, the resulting transcript, and the AI-generated study materials are stored in our database; audio is never used to train external models — we send it through provider APIs configured to opt out of training.
Stage
Provider
Model
Transcription
OpenAI
whisper-large-v3
Reading / summarizing
OpenAI
gpt-4o
Discussion questions
OpenAI
gpt-4o
Quality review (judge)
Anthropic
claude-sonnet-4-6
Scripture embeddings
OpenAI
text-embedding-3-large
Where the AI ends and the pastor begins
Sunday's philosophy: the teaching belongs to the pastor. AI is the packaging.
We use AI to scaffold what the pastor said — extract clips, identify the scripture being preached, write a summary and a study guide, propose discussion questions. We do not use AI as a substitute for the pastor's teaching. The pastor's voice, framing, and theology are the sermon; the AI organizes that material into a form a small group can study during the week.
Concretely, the AI generates the summary, the themes attached to related scripture cards, the clip titles, the discussion questions, and suggested cross-references. The full transcript and the words inside each clip come directly from the audio. Scripture quotations come directly from canonical Bible text.
Before a study guide is published, the pastor can edit any section in the Sunday dashboard.
Accuracy and corrections
Transcription is good but not perfect. If a transcription error makes it to a published study guide, the church is responsible for the fix — either by editing the AI-generated sections that depend on it, or by unpublishing the sermon while the issue is resolved.
Pastors and church admins have full editorial control over everything Sunday generates. The summary, themes, clip titles, discussion questions, and scripture references can all be edited in the dashboard before or after publishing. If the AI got something wrong, the church owns the fix.
If something is wrong, or shouldn't be public
Pastors can unpublish any study guide at any time from the sermon's dashboard page. An unpublished sermon disappears from search engines, scripture pages, sitemaps, and the church's public page within a crawl cycle.
If you spot something wrong on a public study guide and aren't sure who to contact, reach us at hello@sunday.app and we'll route it to the church, or unpublish it ourselves while we do.
What we do not do
We do not publish on our own. Sunday is tooling for churches to publish their own sermons. The church is always the publisher; nothing goes public until they publish it themselves.
We do not train models on sermon audio.
We do not write new theology, new doctrine, or new interpretations. AI-generated content is scaffolding around what the pastor preached, never a replacement for it.