Docs / Troubleshooting

Troubleshooting

Diagnose common environment and generation issues.

Environment

FFmpeg not found

Install the shared build, reopen the terminal, and run diagnostics:

winget install --id Gyan.FFmpeg.Shared -e
llmvoice doctor

The static Gyan.FFmpeg package provides executables but not the shared FFmpeg DLLs required by TorchCodec. Use Gyan.FFmpeg.Shared.

CUDA unavailable

Run:

llmvoice doctor

The installer chooses CUDA only when the supported runtime can be validated. CPU fallback is available but slower. Normal users should rerun the release installer with -Runtime auto or -Runtime cpu rather than manually repairing the isolated environment.

Voice selection

No voice selected

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

Performance

Slow CPU generation

XTTS-v2 is compute intensive. Use --dry-run while validating inputs, close competing workloads, and use NVIDIA CUDA acceleration when the installer reports a supported configuration.

XTTS cannot load

  • Run llmvoice doctor.
  • Allow network access for the first model download.
  • Review and accept the XTTS-v2 CPML terms when prompted.
  • Verify that approximately 4 GB of storage remains available.
  • Retry the command with --debug for technical context.

Files and storage

Output already exists

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

Low disk space

Free space on the application cache and output volumes. Long-form runs need space for model files, temporary chunks, and the final MP3.

Reference audio

Reference rejected

Use a supported audio format with one clear speaker and no music or heavy effects. References must be at least 3 seconds; approximately 6–30 seconds is recommended.

ffprobe reference.wav