Troubleshooting

• CUDA available but model runs on CPU — you installed the CPU-only PyTorch wheel. Reinstall from https://download.pytorch.org/whl/cu121 (or cu118 / cu124 matching your NVIDIA driver).
• flash_attn seems to be not installed — safe to ignore; the backend retries with sdpa.
• Kokoro is silent / no audio — espeak-ng is not on your PATH. Install it (see Installation) and restart the backend.
• Kokoro failed to init for lang_code='j' — install the matching misaki extra: pip install misaki[ja] (or misaki[zh] for Mandarin).
• out of memory during generation — switch to --device cpu or shorten the text. The backend returns a clear error and empties the CUDA cache.
• Switched engines but old cached audio still plays — the cache is per-engine, so old audio stays valid. Click Regenerate on each segment.
• No built-in voices in the sidebar — drop a .wav / .mp3 / .flac / .ogg into backend/voices/ and restart the backend.
• An isolated engine won't install — check the streamed install log in the dialog; each engine builds its own venv on demand and needs a compatible Python (VoxCPM requires Python 3.10–3.12).

Still stuck? Open an issue on GitHub.