Add threading lock around model.generate() to prevent concurrent calls from corrupting shared model state (alignment_stream_analyzer). Update comments to explain the concurrency issue and upgrade path.
feat: add SRT subtitle generation and improve security
- Generate SRT subtitle files alongside WAV audio output by splitting text on sentence boundaries and proportionally distributing timestamps across the audio duration.
- Extend download endpoint to serve both .wav and .srt files with correct MIME types.
- Harden API key authentication: strip whitespace from env var and user input, use `secrets.compare_digest` for constant-time comparison, and reject missing keys.
- Configure Uvicorn with `--forwarded-allow-ips=*` so the Spaces proxy is trusted and `request.base_url` yields https.
- Document upstream constraints (max_new_tokens truncation, watermark) in README and code comments.
Update default model from Chinese variant (v1.1-zh) to base Kokoro-82M and set default voice to `af_heart`. Add voice normalization to handle empty or placeholder "string" values, ensuring robust fallback behavior.