[prosomoalexd]

Picks up #2397 by @kovashikawa, which was closed in favour of #2771 and then approved pending an update to main. This is that work rebased onto main, plus the fixes below. The author's later commit 54dcc0b (whisper.cpp --vad timeline re-anchoring) came after the reviewed head 8763a86, so it's new for review here.

Self-hosters can transcribe recorded videos with their own speech-to-text server instead of AssemblyAI:

STT_BASE_URL=http://whisper:9000/v1
STT_MODEL=large-v3-turbo     # optional, defaults to whisper-1
STT_API_KEY=                 # optional

• With STT_BASE_URL set, recorded-video transcription and editable-transcript backfills use <STT_BASE_URL>/audio/transcriptions (verbose_json, word timestamps). Otherwise AssemblyAI is used as today.
• isTranscriptionConfigured() replaces the ASSEMBLY_API_KEY checks, so an STT endpoint alone is enough. Live transcription of instant recordings stays AssemblyAI-only.
• Changes on top of feat(web): OpenAI-compatible STT endpoint for self-hosted transcription (whisper.cpp, faster-whisper) #2397: AssemblyAI goes through createAssemblyAIClient() (keeps ASSEMBLY_BASE_URL from feat(self-hosting): add ASSEMBLY_BASE_URL for EU transcription #2998); STT_* passed through both compose files (empty values count as unset); a response body cut off after the headers is now retried instead of reported as "did not return JSON"; with AssemblyAI only, the backfill's "no spoken audio" case is no longer retried (it already ended in the same backfill error).
• Docs: "Local Transcription" under AI Features, with whisper.cpp and speaches (as a Compose service) examples. Known limit: with --vad, words right after a long pause can still land early (about 7.5 s early after 8 s of silence in our test).

Coordination: #3520 adds ASSEMBLY_API_KEY guards in four places. Whichever merges second switches them to isTranscriptionConfigured(); STT-only tests here for those four entry points fail on a merge of the two as they stand.

Closes #1996. Refs #2161.

Tests

• 48 new tests: feat(web): OpenAI-compatible STT endpoint for self-hosted transcription (whisper.cpp, faster-whisper) #2397's stt.test.ts and whisper.cpp fixture, plus STT-only cases for every entry point, retry classification (including a real loopback reset and timeout mid-body), and compose's empty values. The STT tests fail on main; the 7 regression controls pass on both.
• Full apps/web suite on Node 24: main 4771 passed / 1 failed, this branch 4819 passed / 1 failed; the same homepage-seo test fails on both. bun run typecheck and biome ci --linter-enabled=false clean.
• Against a local whisper-server (base.en, --convert --vad): mp3 and concatenated segment audio transcribe correctly; silence returns no words.

Confidence Score: 4/5

The PR appears safe to merge, with a non-blocking improvement needed to reject unexpected JSON responses.

Findings

1.  Wrong responses look like silence ▶

Fix with agent prompt

### Issue 1
apps/web/lib/stt.ts:236-241
`parseOpenAICompatibleTranscription` accepts objects without any transcript fields. If a misrouted endpoint returns `{"status":"ok"}`, the parser produces empty words. The workflow then marks the video `NO_AUDIO` and reports success, hiding the configuration error.

Require a recognized transcript shape before accepting an empty result, while keeping valid silent responses supported.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

Adds an OpenAI-compatible speech-to-text provider for recorded videos and editable-transcript backfills.

• Recorded videos can use a self-hosted speech-to-text server.
• Transcription queues and transcript controls recognize local servers.
• Self-hosters can configure the STT server and model in Compose.

_{Reviews (1) · Last reviewed commit: "docs(self-hosting): document local trans..." · Reviewed by Greptile}

[vercel]

@prosomoalexd is attempting to deploy a commit to the Cap Software Inc Team on Vercel.

A member of the Team first needs to authorize it.

[greptile-apps]

Wrong responses look like silence

parseOpenAICompatibleTranscription accepts objects without any transcript fields. If a misrouted endpoint returns {"status":"ok"}, the parser produces empty words. The workflow then marks the video NO_AUDIO and reports success, hiding the configuration error.

Require a recognized transcript shape before accepting an empty result, while keeping valid silent responses supported.

Prompt To Fix With AI

This is a comment left during a code review.
Path: apps/web/lib/stt.ts
Line: 236-241

Comment:
**Wrong responses look like silence**

`parseOpenAICompatibleTranscription` accepts objects without any transcript fields. If a misrouted endpoint returns `{"status":"ok"}`, the parser produces empty words. The workflow then marks the video `NO_AUDIO` and reports success, hiding the configuration error.

Require a recognized transcript shape before accepting an empty result, while keeping valid silent responses supported.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[prosomoalexd]

Fixed in e580f3b: a JSON object with no text, words or segments is now a fatal configuration error instead of silence; {"text":"","segments":[]} still maps to NO_AUDIO.