---

name: clearvoice-inference description: "Operate pretrained ClearVoice inference for speech enhancement, speech separation, speech super-resolution, and audio-visual target speaker extraction." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

ClearVoice Inference

Use this sub-skill when the user wants pretrained ClearVoice inference only.

Route away

• Send SpeechScore or other objective metrics to the sibling speechscore-metrics sub-skill.
• Send training, fine-tuning, config editing, or data generation to the sibling training-and-data-prep sub-skill.

Use it for

• Choosing a ClearVoice task and compatible model_names.
• File, directory, and .scp inference with pretrained checkpoints.
• NumPy/Tensor input mode for audio-only models.
• Output writing, checkpoint download behavior, FFmpeg/media-codec checks, and no-download validation.
• Supported tasks: speech_enhancement, speech_separation, speech_super_resolution, and target_speaker_extraction.

Read or run

• references/api-reference.md when you need the ClearVoice constructor, call, and write() signatures or the tensor/file-mode limits.
• references/model-catalog.md when you need to match a task to the right pretrained model, rate, or checkpoint family.
• references/workflows.md when you need an end-to-end file, directory, .scp, or NumPy workflow and want a safe no-download path first.
• references/troubleshooting.md when inference fails because of unsupported task/model pairs, missing FFmpeg, checkpoint download problems, output conflicts, sample-rate mismatch, or slow CPU fallback.
• scripts/clearvoice_inference_recipe.py when you want a safe CLI recipe that can list models, validate task/model pairs, and run file-mode inference without loading weights in --dry-run.
• scripts/clearvoice_numpy_recipe.py when you want to validate NumPy/Tensor shapes or run a single-model tensor call with --run.

Important rules

• Pass model_names as a list, even for one model.
• Use exactly one model in tensor-to-tensor mode.
• Use video input plus online_write=True for AV target speaker extraction.