[Ahmadnidzam]

Summary

Adds Indonesian (id) support, end to end, for the OmniVoice TTS engine and Whisper STT.

• UI: new id locale (882 keys, same structure as en), registered in the i18n setup and the STT language picker.
• API: id is accepted by the language fields of voice profiles, generation, speak, capture retranscription, and transcription. This is additive.
• Engines: id is listed for OmniVoice. Other TTS engines reject id with HTTP 400 and a clear message. Other language codes are unchanged.
• Robustness: OmniVoice can return an empty or silent waveform for some inputs (a zero-size array error in post-processing). Generation now retries with the seed offset by attempt index, so retries stay deterministic. If every attempt fails, it raises a RuntimeError with a clear message.
• Quality default: NUM_DIFFUSION_STEPS moves from 32 to 64. On a 24-sentence Indonesian benchmark (2 seeds), the median WER on clean sentences went from roughly 0.05-0.07 to 0. Generation takes about twice as long (about 3.8 s vs 1.8 s per clip on an RTX 3050).

Notes for reviewers

• License: OmniVoice weights are CC-BY-NC, as the engine's existing label already says. Indonesian TTS via this path is therefore non-commercial only.
• Language code: OmniVoice's own language map uses id for Indonesian. The override table is unchanged; id passes straight through.
• Behavior change: id requests to engines without Indonesian now return 400. Other languages keep their previous behavior. Only id is guarded, to avoid breaking existing clients.
• Not included: the generated API client under app/src/lib/api is not regenerated. The frontend does not validate language codes through it, so this PR does not need it. Regenerating would touch many unrelated files.
• Latency: the 64-step default doubles OmniVoice generation time.

Testing

• Typecheck (bun run typecheck) passes. bun run build:web passes.
• Backend suite: 342 passed, 11 skipped, 2 failed. Both failures predate this change and come from the Windows environment: test_find_ffmpeg_honours_override_dir expects an extensionless ffmpeg, and test_hf_progress_tracker captures no tqdm updates.
• New tests:
  • test_omnivoice_empty_retry.py (6): retry on empty output, seed offset, final error, instruct errors are not retried, and the degenerate-output check.
  • test_generation_language_guard.py (9): id rejected on engines without it, accepted on OmniVoice, other languages untouched.
• Manual check: 24-sentence Indonesian benchmark with Whisper-medium STT. No TTS failures after the retry fix.

🤖 Generated with Claude Code