talk-cli

talk-cli is a fast, offline-capable multilingual text-to-speech CLI powered by Kokoro. It auto-detects language, supports 150+ voices, and uses the faster MLX backend automatically on Apple Silicon.

talk "Hello, world!"
talk "你好，欢迎使用 talk-cli"
talk -v af_sky -s 1.2 "Speed it up a bit"
talk --output hello.wav "Save to file"

Features

• Auto language detection — switches between English and Chinese based on text content
• 150+ voices — 54 English voices (American, British, Spanish, French, Hindi, Japanese, Portuguese, Italian) and 103 Chinese voices
• Apple Silicon optimised — uses MLX backend for English on M-series Macs, falls back to ONNX automatically
• Offline after first use — engine and models are cached in ~/.cache/talk-cli/
• Save to file — export speech as WAV with --output
• Tiny binary — ~8 MB Go binary; engine and models are downloaded on demand

Supported Platforms

| Platform | Architecture | Backend | |----------|-------------|---------| | macOS | Apple Silicon (arm64) | MLX (fast) + ONNX fallback | | macOS | Intel (amd64) | ONNX | | Windows | amd64 | ONNX | | Linux | — | Export to file only (no audio playback) |

Installation

npm (Node.js 18+)

npm install -g talk-tts

One-line install

macOS:

curl -fsSL https://raw.githubusercontent.com/hoveychen/talk-cli/main/install.sh | bash

Windows (PowerShell):

powershell -c "irm https://raw.githubusercontent.com/hoveychen/talk-cli/main/install.ps1 | iex"

Auto-detects your platform and architecture, installs to ~/.local/bin (macOS) or %LOCALAPPDATA%\talk-cli\bin (Windows), and adds it to your PATH automatically.

To install a specific version or change the install directory:

# macOS
TALK_VERSION=v0.2.0 TALK_INSTALL_DIR=/usr/local/bin curl -fsSL https://raw.githubusercontent.com/hoveychen/talk-cli/main/install.sh | bash

# Windows
$env:TALK_VERSION="v0.2.0"; $env:TALK_INSTALL_DIR="C:\Tools"; powershell -c "irm https://raw.githubusercontent.com/hoveychen/talk-cli/main/install.ps1 | iex"

Build from source

Requires Go 1.22+.

git clone https://github.com/hoveychen/talk-cli.git
cd talk-cli
make build       # outputs to bin/

Usage

Speak text

talk "Hello, world!"                        # English (auto-detected)
talk "你好，欢迎使用 talk-cli"               # Chinese (auto-detected)
talk --lang zh "Kokoro is great"            # Force language
talk -v af_sky "Choose a specific voice"    # Choose voice
talk -s 1.5 "Speak 50% faster"             # Adjust speed (0.5–2.0)
talk -o out.wav "Save as WAV"              # Export to file

List voices

talk voices              # All voices
talk voices --lang en    # English only
talk voices --lang zh    # Chinese only

Pre-download assets (for offline use)

talk init                # Download both English and Chinese
talk init --lang en      # English only
talk init --lang zh      # Chinese only

All flags

Flags:
      --lang string     Language: auto, en, zh (default "auto")
  -v, --voice string    Voice name (default depends on language)
  -s, --speed float     Speed multiplier 0.5–2.0 (default 1.0)
  -o, --output string   Save WAV to file instead of playing
      --no-progress     Suppress download progress bar
  -h, --help            Help

How it works

On first use, talk downloads the appropriate engine bundle and model:

~/.cache/talk-cli/
├── en/
│   ├── engine/     # PyInstaller-packaged Kokoro ONNX or MLX engine
│   └── model/      # model.onnx + voices.bin (~88 MB INT8)
└── zh/
    ├── engine/     # PyInstaller-packaged Kokoro ONNX engine
    └── model/      # model.onnx + voices.bin + config.json (~82 MB INT8)

The Go binary manages downloads, invokes the engine subprocess with JSON arguments, and plays back the WAV output. Language auto-detection uses Unicode CJK block (U+4E00–U+9FFF).

Models

| Language | Model | Size | Source | |----------|-------|------|--------| | English | Kokoro v1.0 (INT8 ONNX) | ~88 MB | thewh1teagle/kokoro-onnx | | Chinese | Kokoro v1.1-zh (INT8 ONNX) | ~82 MB | hoveyc/talk-cli-models | | English (MLX) | Kokoro-82M-bf16 | streamed | mlx-community/Kokoro-82M-bf16 |

AI Agent Integration

talk-cli ships with AI skills compatible with Claude Code, Cursor, and other AI coding agents. Install the skill to let your AI assistant use talk-cli:

npx skills add hoveychen/talk-cli

Contributing

Contributions are welcome! See CONTRIBUTING.md for guidelines.

License

MIT