VoxCPM CLI is a slim local command-line tool for VoxCPM2 text-to-speech, running entirely on your machine through OpenVINO — no server, no PyTorch required at runtime. It ships a JSON-first CLI, machine-friendly exit codes, and an agent Skill, so it works equally well from a terminal, a script, or an LLM agent.

Why

• Local & private — speech is generated on your machine; nothing is sent to a third-party API.
• Agent-friendly — every command emits structured JSON on stdout and human logs on stderr, with meaningful exit codes.
• Inference without PyTorch — the model is converted to OpenVINO IR, so synthesis runs on a lightweight OpenVINO runtime (CPU or GPU) instead of a full PyTorch stack.
• Slim repo — model weights and converted artifacts are not stored in git; prepare downloads and converts them explicitly when you approve.

How it works

Plain Text

┌────────────┐  download/convert  ┌───────────────────────┐
│ HF + GitHub ──────────────────► │ cache/  models/       │
└────────────┘                     │ (gitignored artifacts)│
                                   └───────────┬───────────┘
                                               │
                          voxcpm              │ synth
                                               ▼
                                   ┌───────────────────────┐
                                   │ OpenVINO pipeline      │
                                   │ (no PyTorch at runtime)│
                                   └───────────┬───────────┘
                                               ▼
                                   output/*.wav  +  JSON result

prepare downloads the upstream VoxCPM source and VoxCPM2 weights, then converts them into OpenVINO IR. synth loads the IR artifacts with OpenVINO and writes a 48 kHz WAV to output/.

Requirements

Item	Requirement
OS	Windows
Python	3.10, 3.11, or 3.12
Network	First prepare run only (model download + source download)
Disk	Enough for upstream source, original weights, and OpenVINO IR

Runtime dependencies (declared in pyproject.toml): huggingface-hub, librosa, numpy, openvino, soundfile, tokenizers.

Conversion-only extras, needed when running prepare from a slim checkout: nncf, safetensors, torch>=2.5.0, transformers>=4.36.2.

Installation

From the project root:

Bash

python -m pip install -e .

To also be able to download and convert the model locally:

Bash

python -m pip install -e ".[convert]"

Quick start

Bash

# 1. Check whether models and OpenVINO are ready (never downloads anything)
voxcpm status --json

# 2. If not ready, download & convert (network + disk use — approve first)
voxcpm prepare --json

# 3. Synthesize
voxcpm synth --text "你好" --json

pip install registers the voxcpm console script; python -m voxcpm_cli also works when the package is on the path.

CLI reference

All commands accept --json for JSON output on stdout. Progress and logs go to stderr, so stdout stays machine-parseable.

status

Report whether the expected artifacts exist and which OpenVINO device would be used:

Bash

voxcpm status --json

Expected directories:

Plain Text

cache/source/VoxCPM/
models/original/VoxCPM2/
models/openvino/VoxCPM2/

prepare

Download the upstream source, the VoxCPM2 weights (Hugging Face), and convert them into OpenVINO IR under models/openvino/VoxCPM2/:

Bash

voxcpm prepare --json

Useful options:

Bash

voxcpm prepare --json --device AUTO
voxcpm prepare --json --force-convert
voxcpm prepare --json \
  --model-dir models/original/VoxCPM2 \
  --ov-model-dir models/openvino/VoxCPM2

synth

Generate a WAV file. Provide exactly one of --text or --text-file:

Bash

# Short text
voxcpm synth --text "你好" --json

# Long text from a file
voxcpm synth --text-file input.txt --json

# Explicit output path (must stay inside output/)
voxcpm synth --text-file input.txt --output output/demo.wav --json

# Voice instruction
voxcpm synth --text-file input.txt \
  --voice-instruction "温柔、自然、语速适中" --json

# Generation parameters
voxcpm synth --text-file input.txt \
  --cfg-value 2.0 --inference-timesteps 10 --max-len 2000 \
  --device AUTO --json

On success, stdout is JSON:

JSON

{
  "ok": true,
  "path": "C:\\project\\voxcpm-cli\\output\\demo.wav",
  "sample_rate": 48000,
  "format": "wav",
  "model": "VoxCPM2 OpenVINO",
  "device": "GPU",
  "duration_ms": 12345
}

--version

Bash

voxcpm --version

Error handling

Failures are JSON too, and each maps to a process exit code:

JSON

{
  "ok": false,
  "error": { "code": "MODEL_NOT_READY", "message": "Run prepare before synth." }
}

Exit code	Meaning
0	success
1	validation error
2	model prepare/load error
3	synthesis error
4	file write error

Common error codes:

• NO_TEXT_INPUT
• INVALID_ARGUMENT
• INVALID_OUTPUT_PATH
• MODEL_NOT_READY
• VOXCPM_SOURCE_DOWNLOAD_FAILED
• MODEL_DOWNLOAD_FAILED
• MODEL_CONVERSION_FAILED
• MODEL_LOAD_FAILED
• OPENVINO_DEVICE_UNAVAILABLE
• TTS_GENERATION_FAILED
• AUDIO_WRITE_FAILED

Using an agent

The repo includes an agent Skill in skills/voxcpm-tts/SKILL.md. The workflow it describes:

1. Run voxcpm status --json.
2. If not ready, ask the user before running voxcpm prepare --json.
3. Write long text to a temporary .txt file.
4. Run voxcpm synth --text-file <file> --json.
5. Return the WAV path, sample rate, and format.

Rules: use the local CLI (never a server); do not expose sensitive text in final replies; do not download or convert models without explicit approval.

Project structure

Plain Text

voxcpm-cli/
├── pyproject.toml            # Package + dependencies (and the convert extra)
├── LICENSE                   # Apache-2.0
├── assets/logo.svg           # README header logo
├── README.md
├── voxcpm_cli/
│   ├── __main__.py          # voxcpm entry
│   ├── cli.py               # argparse + JSON output, exit codes
│   ├── engine.py            # status / prepare / synthesize orchestration
│   ├── paths.py             # Artifact & output-path resolution
│   └── voxcpm2_tts_helper.py# Conversion + OpenVINO inference pipeline
├── skills/voxcpm-tts/       # Agent Skill instructions
├── cache/                   # Downloaded upstream source (gitignored)
├── models/original/         # Downloaded weights (gitignored)
├── models/openvino/         # Converted OpenVINO IR (gitignored)
└── output/                  # Generated WAVs (gitignored)

Licensed under Apache-2.0.

Links

• GitHub: wuhaostudio/voxcpm-cli
• Agent Skill: skills/voxcpm-tts/SKILL.md
• Author: wuhaostudio