fix(transcription): repair AssemblyAI deprecation, add providers, harden upload/YouTube crashes

[steethebutcher]

Summary

Fixes transcription crashes from the deprecated AssemblyAI speech_model parameter, integrates multiple transcription providers, hardens several YouTube/upload crash paths, and raises the local-upload size limit so large files actually upload.

Background & PR review

Reviewed the open PRs targeting the transcription/AssemblyAI area. Several attempt the same fix but none merge cleanly (all target a pre-#63 base before the _assemblyai_speech_model_value refactor), so the logic was ported onto current main:

• fix(transcription): switch to speech_models list for AssemblyAI Asse… #47 (AssemblyAI speech_models) → integrated (adapted)
• feat(transcription): add support for multiple transcription providers and update configuration #55 (multi-provider transcription) → integrated (adapted)
• Fix Postgres connection-pool exhaustion in clip file/export routes #50 (Postgres connection-pool exhaustion) → integrated (direct port)
• fix: Replace deprecated speech_model with speech_models parameter #43, fix: speech_model is deprecated use speech_models instead, change config variable to app_config (variable name conflict between library) and change result_type to output_type #17 → skipped (superseded / touched unrelated pydantic-ai output_type/result_type/app_config params)
• feat: add faster-whisper and local Whisper support #53, feat: Local Transcription and image detection #60 → deferred (too invasive / oversized)
• feat(backend): hybrid face tracking, AssemblyAI SDK fix, and runtime bug fixes #30, fix: apply PR#30 patches for upload bug and AssemblyAI SDK compatibility #35 → skipped (the cleanup_settings "crash" they fix is already fixed on main)

Changes

1. AssemblyAI deprecation fix (video_utils.py)

AssemblyAI deprecated the singular speech_model (best/nano/universal) server-side; requests now require the plural speech_models list (universal-3-pro, universal-2). Replaced _assemblyai_speech_model_value → _assemblyai_speech_models_value, mapping legacy aliases: nano/universal-2 → ["universal-2"]; everything else → ["universal-3-pro", "universal-2"].

2. Transcription provider choice (TRANSCRIPTION_PROVIDER)

New config: assemblyai (default) | whisper | youtube_captions.

• Local Whisper with word-level timestamps; Whisper seconds are converted to ms to match the transcript-cache units so subtitles align.
• YouTube-captions extraction via yt-dlp.
• cache_transcript_data, format_transcript_for_analysis, _serialize_transcript_word are now polymorphic (AssemblyAI objects + Whisper dicts).
• generate_transcript now threads source_url so the youtube_captions provider works.

3. Local upload size limit (the large-upload bug)

The frontend cap was raised to 12 GB but the backend still hardcoded 1 GB, so large uploads hit POST /upload and were rejected with HTTP 413. Made the limit configurable (MAX_VIDEO_UPLOAD_BYTES, default 12 GB) and fixed stale "under 1 GB" client messages.

4. Connection-pool exhaustion (#50)

get_clip_file / export_clip held the Depends(get_db) session open for the whole FileResponse/ffmpeg export → "idle in transaction" → QueuePool exhaustion under normal playback. Added await db.close() after the clip lookup in both routes, plus an idle_in_transaction_session_timeout safety net on the engine.

5. YouTube/download crash hardening

• Downloaded source videos were never cleaned up → disk fills → silent crashes. Now cleaned on both success and error paths (uses the previously-dead cleanup_downloaded_files).
• Tasks that died mid-"processing" were stuck forever (only "queued" stale tasks were recovered). Added _is_stale_processing_task + PROCESSING_TASK_TIMEOUT_SECONDS (default 4h, safely above arq's 3h job_timeout).

Env / config impact

New optional env vars (all have safe defaults, backward compatible):

• TRANSCRIPTION_PROVIDER (default assemblyai)
• MAX_VIDEO_UPLOAD_BYTES (default 12000000000)
• PROCESSING_TASK_TIMEOUT_SECONDS (default 14400)

No DB migrations.

Verification

• python -m compileall src clean; all edited modules import without cycles.
• Backend tests: 103 passed, 9 skipped. Added 4 new tests (AssemblyAI speech_models mapping, Whisper provider dispatch, Whisper-dict formatting & caching).
• The 1 failing test (test_burn_ass_subtitles_passes_selected_fonts_dir) is pre-existing on clean main — a Windows path-escaping artifact in an ffmpeg filter, unrelated to this change.

Manual verification steps

• TRANSCRIPTION_PROVIDER=assemblyai: confirm speech_models request succeeds (no more deprecation 400).
• TRANSCRIPTION_PROVIDER=whisper: short video transcribes locally, subtitles align.
• Upload a >1 GB file: succeeds (was: HTTP 413).
• Play/export a clip under concurrent load: no "idle in transaction" pile-up.

[vercel]

@steethebutcher is attempting to deploy a commit to the Shiori LLC Team on Vercel.

A member of the Team first needs to authorize it.