ClearerVoice-Studio

Use this repo skill when a task involves ClearerVoice-Studio, the clearvoice package, ClearVoice pretrained speech processing models, SpeechScore quality metrics, or the repository's training and data-preparation workflows.

Fast route map

User task	Read
Use pretrained models for speech enhancement, speech separation, speech super-resolution, or audio-visual target speaker extraction	sub-skills/clearvoice-inference/SKILL.md
Choose task/model_names, validate file/directory/.scp inputs, handle NumPy/Tensor inference, or debug checkpoint/media issues	sub-skills/clearvoice-inference/SKILL.md
Compute PESQ, STOI, SI-SDR, SNR, DNSMOS, NISQA, DISTILL_MOS, or other SpeechScore metrics	sub-skills/speechscore-metrics/SKILL.md
Score matched directories, decide whether a clean reference is required, or troubleshoot metric dependencies	sub-skills/speechscore-metrics/SKILL.md
Train, fine-tune, resume, run training-side inference, prepare .scp/CSV data lists, or generate noisy/reverb training data	sub-skills/training-and-data-prep/SKILL.md
Check whether a local environment is ready without downloading weights or launching training	scripts/check_clearer_voice_environment.py

Component overview

Read references/package-and-component-overview.md when you need the repository component map, verified public signatures, package/source layout notes, or high-level prerequisites.

Key facts:

• clearvoice is the installable package for ClearVoice inference: from clearvoice import ClearVoice.
• The public ClearVoice constructor is ClearVoice(task, model_names).
• Calling a ClearVoice instance with a string path uses file/directory/.scp I/O mode; calling with a NumPy array or Torch tensor uses tensor-to-tensor mode.
• SpeechScore is a source-layout component in this snapshot; use its helper or source-layout notes when importing it outside its component directory.
• Training workflows are script/config based and should be treated as templates that require user-owned datasets, checkpoints, GPUs, and explicit output paths.

Safe first checks

Before running expensive or mutating workflows:

1. Run python scripts/check_clearer_voice_environment.py to check clearvoice, Torch/CUDA visibility, and FFmpeg presence without loading model weights.
2. For SpeechScore source-layout checks, pass the user's SpeechScore component directory to --speechscore-dir.
3. Use each sub-skill's dry-run helper before real model inference, metric scoring, or training launch preparation.
4. Ask before starting model downloads, distributed training/evaluation, data generation, checkpoint overwrites, or package backend changes in a user-owned environment.

Installation guidance

For pretrained ClearVoice package use:

pip install clearvoice
python - <<'PY'
from clearvoice import ClearVoice
print(ClearVoice)
PY

For full repository workflows such as SpeechScore metrics and training scripts, install the repository's documented runtime requirements in an isolated environment. Do not modify a user's existing working environment without approval.

Troubleshooting

• Read references/troubleshooting.md for cross-cutting install/import, CUDA, FFmpeg, model-download, and component-routing failures.
• Use sub-skill troubleshooting references for workflow-specific symptoms and recovery steps.
• Read references/repo-provenance.md before deciding whether this skill is stale for a different checkout or package version.

Boundaries

Use this skill to operate the package/repository workflows, not to modify the ClearerVoice-Studio source code. If the user asks for generic speech recognition, TTS, diarization, or unrelated PyTorch training, route to a more specific skill. If the user asks to edit this repository's source, treat it as a repository-maintenance task and inspect the current checkout directly.