tylerbcrawford/subgeneratorr v3.2.0

v3.2.0 — current LLM models, Deepgram SDK 7, four new languages (as, kk, mn, ps), 58 translation targets

on GitHub

This release brings every external dependency up to date: current LLMs for keyterms and translation, Deepgram SDK 7, and four new Nova-3 languages. It also fixes a batch of reliability issues found by a full end-to-end test run.

Current LLM models

• The model list lives in one place (core/llm_providers.py), so the Web UI dropdowns, cost estimates, and backend stay in sync.
• Each model is pinned to its lowest thinking/reasoning setting, which keeps keyterm and translation calls fast and cheap.
• Retired model ids (claude-sonnet-4-6, gpt-4.1, gemini-2.5-flash, …) are still accepted and map to their replacements, so saved settings keep working.

Benchmarks for all six models (keyterm quality, translation speed, and cost) are in docs/technical.md.

Deepgram

• deepgram-sdk 7.x and deepgram-captions 2.x. Caption output is byte-identical to 1.x for the same response.
• Speaker diarization v2 in the CLI too. The CLI was still on Deepgram's legacy v1 diarizer.
• Four new transcription languages: Assamese (as), Kazakh (kk), Mongolian (mn), and Pashto (ps). Each writes its own tagged sidecar (.asm.srt, .kaz.srt, .mon.srt, .pus.srt) and is also a translation target, for 58 in total.
• Removed English (Ireland) and English (Canada). Deepgram rejects both with HTTP 400. Every remaining language code was verified against the live API.
• Cost estimates follow Deepgram's published rates and your options: $0.0043/min base, $0.0052/min for multi, +$0.0013/min with keyterms, +$0.0020/min with redaction. Estimates previously used a flat $0.0057/min.
• New DEEPGRAM_TIMEOUT (default 900 s). The SDK's own 60 s default is too short for feature-length audio.

Fixes

• Speaker transcripts had no punctuation or capitalization. They now match the SRT (Web and CLI).
• Ollama on slow hardware: timed-out local generations were silently retried twice, burning up to ~30 minutes of CPU per window. They now fail fast with an actionable message. Set OLLAMA_TIMEOUT (default 600 s) for big models on slow CPUs.
• Keyterm generation without an explicit model used an invalid default. Unknown model ids now return HTTP 400 instead of failing inside the worker.
• Claude responses that begin with a thinking block, plus refusals and max-token truncation, are handled for every provider.
• The chunk-confirmation dialog no longer prices local Whisper batches as Deepgram.
• Docs: Ollama is a translation-only provider (keyterms can still be entered manually), and the keyterm-generation API docs now describe its async task_id flow.

Upgrading

Pull the new images (:3.2.0 or :latest) and recreate the web and worker containers. No configuration changes are required. If you use the example compose files, copy in the new optional DEEPGRAM_TIMEOUT and OLLAMA_TIMEOUT lines, because they forward env vars explicitly.

See CHANGELOG.md for the complete list.