[vincentfranstyo]

Problem

Generating with LuxTTS on a short input like "This is a test" fails every time with a raw torch error shown in the UI:

RuntimeError: Calculated padded input size per channel: (6). Kernel size: (7). Kernel size can't be greater than actual input size

The text produces too few feature frames for the Vocos vocoder's Conv1d embed layer (kernel size 7).

Fix

In LuxTTSBackend.generate(), catch that specific RuntimeError around generate_speech() and raise a ValueError with a clear message ("Text is too short for LuxTTS. Try a longer sentence or another engine."). The generation service already stores str(e) as the generation error, so this message is what the user sees. Any other RuntimeError is re-raised unchanged.

I went with catching the error rather than a fixed minimum character count, because the frame count depends on the text and the voice prompt, so a hard-coded threshold would either reject valid input or miss some crashes.

Testing

• Added backend/tests/test_luxtts_short_text.py with a fake model: the kernel-size error becomes a ValueError, and an unrelated RuntimeError still propagates. The first test fails on main and passes with this change.
• ruff check / ruff format --check clean on the new test.
• Not tested with the real LuxTTS weights.

Fixes #1046