Docs / Troubleshooting

Troubleshooting

Diagnose common environment and generation issues.

Environment

FFmpeg not found

Confirm ffmpeg -version works in the same terminal where LLMVoice runs. Restart the terminal after updating PATH, then rerun llmvoice doctor.

CUDA not found

Check the NVIDIA driver and verify that your installed PyTorch build reports CUDA support. Use CPU explicitly if you need a reliable fallback:

llmvoice config set device cpu

Voice selection

No voice selected

Pass a voice for the current run or set a default:

llmvoice start transcript.txt --voice friday
llmvoice config set default_voice friday

Use llmvoice voice list to verify the name.

Performance

Slow CPU generation

XTTS-v2 synthesis is compute intensive. Reduce the transcript while testing, use --dry-run for validation, close competing workloads, or move to a supported CUDA device.

Files and storage

Output already exists

Choose a different output path or deliberately allow replacement with --force.

Low disk space

Free space on the model-cache, temporary, and output volumes. Long-form runs need room for intermediate chunks as well as the final MP3.

Reference audio

Invalid reference audio

Open the file in an audio player, confirm it is not corrupted, and convert it to a standard PCM WAV with FFmpeg:

ffmpeg -i reference-input.m4a -ac 1 -ar 44100 reference.wav