Add Qwen (Alibaba) qwen-audio-3.0-tts as a fourth TTS backend

[ballPointPenguin]

What & why

Adds a fourth TTS backend, qwen (Alibaba Model Studio's newly released qwen-audio-3.0-tts), alongside ElevenLabs, Mistral (Voxtral), and xAI (Grok). It follows the xAI backend pattern end to end so professors can be assigned a Qwen preset voice from the voice-selection UI.

Changes

Backend

• New artificial_u/integrations/qwen/ package:
  • QwenTTSClient — wraps the official dashscope SDK SpeechSynthesizer (the 3.0 models are WebSocket-only). Requests MP3 output so TTSService's existing chunk-concat + ffmpeg remux pipeline works unchanged. Retry/backoff mirrors the xAI client. Resolves flash vs. plus model from the chosen voice.
  • QwenVoiceManager — static 14-voice preset catalog (Alibaba exposes no voice-list API for these models).
• QwenTTSBackend adapter, registered in integrations/tts/factory.py.
• ALIBABA_API_KEY, TTS_QWEN_MODEL, ALIBABA_TTS_WSS_URL in settings/defaults, wired through system.py. Explicit qwen branch in the speech processor.

API + frontend

• GET /api/v1/voices/qwen/catalog (static list, filterable by language/gender).
• qwen enrichment branch in VoiceService.manual_voice_assignment_generic — extracted per-provider enrichment into a _enrich_generic_voice helper to stay under the flake8 complexity limit. Preview and assign already work generically.
• New Qwen (Alibaba) tab in ProfessorVoice.tsx with a QwenVoiceCard (name, description, gender, flash/plus tier badge), plus types.ts/config.ts/voice-service.ts and i18n keys in en/es/fr/zh.

Tests / CI / docs

• New tests/unit/integrations/test_qwen_tts.py (13 tests), api-key assertions in test_api_keys.py + conftest.py, ALIBABA_API_KEY added to both CI workflows, speech-processor test, scripts/seed_qwen_voices.py, and docs/configuration.md.

Verification

• All 344 unit tests pass; backend black/isort/flake8 and frontend eslint/biome/stylelint/build all clean.
• Catalog endpoint returns correct data with language/gender filters.
• Live smoke test confirms the integration is correct up to the auth boundary (well-formed SDK call, model, voice, MP3 request).

⚠️ Reviewer note — region availability

qwen-audio-3.0-tts is hosted only in the Singapore and Beijing regions, not us-east-1 (Virginia). Verified by probing with a real key:

Region	HTTP text-gen	WebSocket TTS
us-east-1 (dashscope-us)	200 (key valid)	404 — no websocket TTS route
Singapore (dashscope-intl)	401	401 InvalidApiKey — route/model exist, key rejected
Beijing (dashscope)	—	401 InvalidApiKey

Alibaba API keys are region-locked, so a Singapore- (or Beijing-) region ALIBABA_API_KEY is required to generate audio. The default ALIBABA_TTS_WSS_URL is set to Singapore (dashscope-intl) accordingly — a us-east-1 default provably cannot serve this model.

Deferred

• Voice cloning ("clone-to-qwen") and Voice Design — the international qwen-voice-enrollment cloning API currently targets only the qwen3-tts-vc family, not 3.0, and Voice Design has no international API docs yet. A clone-to-qwen flow can later mirror the existing clone-to-mistral feature.

🤖 Generated with Claude Code