feat: chunked TTS generation with quality selector

[glaucusj-sai]

Summary

Long text that exceeds the Qwen3-TTS model's max_new_tokens=2048 limit (~170s audio) now gets automatically handled:

• Text splitting: Splits at sentence boundaries (with clause/word fallbacks) into configurable chunks (default 800 chars)
• Crossfade concatenation: Joins audio chunks with a 50ms crossfade to eliminate clicks at boundaries
• Quality selector: Runtime-switchable between standard (24kHz native) and high (44.1kHz via soxr VHQ resampling)
• Settings API: New GET/POST /tts/settings endpoints for runtime quality control without restart

Short text (<800 chars) uses the original single-shot fast path with zero overhead.

Changes

File	Change
backend/utils/chunked_tts.py	New: text splitting, audio concat, resampling utilities
backend/backends/pytorch_backend.py	Integrate chunking into generate(), extract _generate_single()
backend/main.py	Add GET/POST /tts/settings endpoints
backend/models.py	Add TTSSettingsUpdate model, bump text max_length to 50000
backend/requirements.txt	Add soxr>=0.3.0 for high-quality resampling

Environment variables

Variable	Default	Description
TTS_QUALITY	standard	Output quality (standard=24kHz, high=44.1kHz)
TTS_MAX_CHUNK_CHARS	800	Max characters per chunk
TTS_UPSAMPLE_RATE	44100	Target sample rate for high quality

Test plan

• Short text (<800 chars): uses single-shot path, no chunking overhead
• Long text (9K+ chars): splits into ~12 chunks, generates and concatenates seamlessly
• Quality switch to high: output sample rate changes to 44100
• Switch back to standard: output returns to 24000
• GET /tts/settings returns current config
• POST /tts/settings with {"quality":"high"} updates at runtime

Tested on NVIDIA DGX Spark with Qwen3-TTS 1.7B — 9K character input produced ~12 minutes of seamless audio.

[TacoDark]

AI Generated pull request. Please review code to make sure it works.

[glaucusj-sai]

AI Generated pull request. Please review code to make sure it works.

yes it has been tested and i have used AI to post as i didn't have much experence with forking. this was done for my project that needed large scripts for podcast where quality matters and need long text to convert in the voice. I have generated more then 10 hours of voice so far with this. thanks for your commets.

[TacoDark]

AI Generated pull request. Please review code to make sure it works.

yes it has been tested and i have used AI to post as i didn't have much experence with forking. this was done for my project that needed large scripts for podcast where quality matters and need long text to convert in the voice. I have generated more then 10 hours of voice so far with this. thanks for your commets.

Thank you for being honest, I approve of commit