@acpfx/tts-pocket
v0.2.8
Published
Local TTS via Pocket TTS (on-device, Rust/Candle, CPU-capable)
Readme
@acpfx/tts-pocket
Local text-to-speech via Pocket TTS. A lightweight ~100M parameter model that runs on-device, including on CPU -- no API key needed.
Usage
This package is a pipeline node for @acpfx/cli. See the CLI package for installation and usage.
The postinstall script downloads a prebuilt binary for your platform. On Linux/Windows with an NVIDIA GPU (Ampere or newer, compute capability 8.0+), a CUDA-accelerated binary is downloaded automatically. Otherwise falls back to CPU (~6x realtime on Apple Silicon).
Manifest
- Consumes:
agent.delta,agent.complete,agent.tool_start,control.interrupt - Emits:
audio.chunk,lifecycle.ready,lifecycle.done,log
Settings
| Name | Type | Default | Description |
|------|------|---------|-------------|
| voice | string | | Voice name or path to WAV/safetensors file for voice cloning |
| temperature | number | 0.7 | Generation temperature (0 = deterministic) |
| variant | string | b6369a24 | Model variant/config name |
Pipeline Example
nodes:
tts:
use: "@acpfx/tts-pocket"
settings: { temperature: 0.7 }
outputs: [player]GPU Acceleration
| Platform | Acceleration | Requirement | |----------|-------------|-------------| | macOS | Metal | Apple Silicon (automatic) | | Linux/Windows | CUDA | NVIDIA Ampere+ (RTX 3090, A100, etc.) | | All | CPU | Fallback, ~6x realtime on Apple Silicon |
Building from Source
cargo build --release -p node-tts-pocket
# With Metal (macOS):
cargo build --release -p node-tts-pocket --features metal
# With CUDA (Linux/Windows, requires CUDA toolkit):
cargo build --release -p node-tts-pocket --features cudaExternal Links
- Pocket TTS (Kyutai) -- Original Pocket TTS model
- Pocket TTS (Rust port) -- Rust/Candle implementation used by this node
License
ISC
