Local TTS — Offline Text-to-Speech

Generate speech locally with VoxCPM2. 30 languages, voice design, voice cloning. Zero cloud, zero cost.

Runs 100% on your machine using VoxCPM2 (2B-parameter model, Apache-2.0). Optimized for Apple Silicon via Metal (MPS). No API keys, no rate limits, no telemetry.

Features

| Feature | Description | |---------|-------------| | Text-to-Speech | 30 languages, auto-detected from input text | | Voice Design | Describe a voice in natural language (e.g. "warm female voice, mid-30s") | | Voice Cloning | Clone any voice from a 3-10 second reference clip | | Ultimate Cloning | Reference + prompt for maximum fidelity (vocal micro-nuances) | | 48 kHz output | Production-quality WAV ready for Telegram, video, podcast |

Installation

/plugin install local-tts@claude-code-plugins-plus

Requirements

• Python 3.10+ (3.12 recommended)
• macOS with Apple Silicon (M1/M2/M3/M4) — uses Metal for acceleration. Linux with CUDA also supported but not the default target.
• ~10 GB disk for model weights (downloaded on first use to ~/.cache/huggingface/)
• ~16 GB RAM recommended

First-run setup

The plugin uses a dedicated Python venv. Create it once:

mkdir -p ~/.local-tts
python3.12 -m venv ~/.local-tts/venv
~/.local-tts/venv/bin/pip install --upgrade pip
~/.local-tts/venv/bin/pip install voxcpm

On first call, VoxCPM2 downloads ~10 GB of model weights. Subsequent calls reuse the cache (30s startup vs 5+ min first time).

Usage

Via natural language (auto-triggered)

Just ask Claude to:

• "Say hello in French"
• "Generate a voiceover for this text: ..."
• "Clone this voice: /path/to/sample.wav and say ..."
• "Make a warm female voice reading: ..."

Direct invocation

VENV=~/.local-tts/venv
SCRIPT=${CLAUDE_PLUGIN_ROOT}/scripts/generate.py

# 1. Basic TTS (default voice)
"$VENV/bin/python" "$SCRIPT" --text "Hello world" --out /tmp/hello.wav

# 2. Voice Design (no reference needed)
"$VENV/bin/python" "$SCRIPT" \
  --text "(warm female voice, mid-30s, calm)Welcome back." \
  --out /tmp/design.wav

# 3. Voice Cloning (3-10s reference)
"$VENV/bin/python" "$SCRIPT" \
  --text "This is my cloned voice." \
  --ref /path/to/sample.wav \
  --out /tmp/clone.wav

# 4. Long text via stdin
cat article.txt | "$VENV/bin/python" "$SCRIPT" --stdin --out /tmp/article.wav

Supported languages (30)

Arabic, Burmese, Chinese (+ dialects), Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Indonesian, Italian, Japanese, Khmer, Korean, Lao, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swahili, Swedish, Tagalog, Thai, Turkish, Vietnamese.

No language tag needed — VoxCPM auto-detects from the text.

Performance (Apple M4, MPS, bfloat16)

| Metric | Value | |--------|-------| | First model load | ~340s (downloads 10 GB weights) | | Subsequent loads | ~30s | | Generation RTF | ~2.3× realtime (10s audio ≈ 23s compute) |

Not suitable for real-time streaming. Great for batch generation, voiceovers, podcasts, Telegram voice messages.

Script options

| Flag | Description | |------|-------------| | --text STR | Text to synthesize | | --stdin | Read text from stdin (for long input) | | --out PATH | Output WAV path (required) | | --ref PATH | Reference audio for cloning | | --prompt-wav PATH | Prompt wav for ultimate cloning (max fidelity) | | --cfg FLOAT | Classifier-free guidance (default 2.0) | | --steps INT | Diffusion steps (default 10, higher = slower but marginally better) | | --model ID | Model id (default openbmb/VoxCPM2) | | --quiet | Suppress loading messages |

Troubleshooting

ModuleNotFoundError: voxcpm — venv missing or wrong path. Run the setup commands above.

No such file: VoxCPM2 weights — HuggingFace cache missing. First run will download (needs network).

Slow first call — normal, model load takes 30s after cache warms. Subsequent calls in the same Python process are instant; this script spawns a fresh process per call. For batch work, write a wrapper that loads the model once.

French pronunciation of names — add an IPA-ish hint or rephrase. Most names work out of the box.

Links

• VoxCPM2 repo: https://github.com/OpenBMB/VoxCPM
• Model card: https://huggingface.co/openbmb/VoxCPM2
• Source plugin: https://github.com/vdk888/local-tts (same codebase)

License

Apache-2.0 — Bubble Invest 2026