ltts

Quick CLI for local text-to-speech with two backends: Qwen3-TTS (default) and Kokoro TTS.

Install

Recommended (fast, reproducible):

uv tool install ltts

Run without installing:

uvx ltts "hello world" --say

With pip:

pip install ltts

NVIDIA GPU (Optional)

For faster inference on NVIDIA GPUs:

pip install 'ltts[cuda]'

Usage

# Generate speech (saves to output.mp3 by default)
ltts "Hello, world!"

# Play through speakers
ltts "Hello, world!" --say

# Save to specific file
ltts "Hello, world!" -o speech.wav

# Read from stdin
echo "Hello from pipe" | ltts --say
cat article.txt | ltts -o article.mp3

Backends

Qwen3-TTS (default)

Higher quality with voice cloning and emotional control. Supports 10 languages.

# Preset voices
ltts "Hello, world!" -v Ryan --say       # English male (default)
ltts "Hello, world!" -v Aiden --say      # English male
ltts "你好世界" -v Vivian --say           # Chinese female
ltts "こんにちは" -v Ono_Anna --say       # Japanese female
ltts "안녕하세요" -v Sohee --say           # Korean female

# Voice cloning (3+ seconds of reference audio)
ltts "Hello in your voice" --ref-audio voice.wav --say
ltts "Hello" --ref-audio voice.wav --ref-text "transcript" --say

# Emotional control
ltts "I can't believe we won!" --instruct "speak with excitement" --say

# Smaller model for faster inference
ltts "Hello world" --model-size 0.6B --say

Preset voices: Ryan, Aiden (English), Vivian, Serena, Dylan, Eric, Uncle_Fu (Chinese), Ono_Anna (Japanese), Sohee (Korean)

Languages: en, zh, ja, ko, de, fr, es, pt, it, ru

Kokoro TTS

Lightweight with 50+ voices. Supports streaming for faster time-to-first-audio.

# Use Kokoro backend
ltts "Hello world" -b kokoro -v af_heart --say
ltts "こんにちは" -b kokoro -v jf_alpha --say

# Stream chunks as generated (lower latency)
ltts "Hello world" -b kokoro --say --chunk

Voices: af_heart, af_alloy, af_bella, am_adam, am_michael (American), bf_alice, bf_emma, bm_daniel (British), jf_alpha, jm_kumo (Japanese), zf_xiaobei, zm_yunxi (Chinese), ef_dora, em_alex (Spanish), ff_siwis (French), and more.

Full voice list: https://huggingface.co/hexgrad/Kokoro-82M/blob/main/VOICES.md

Options

# Device selection
ltts "Hello" -d cpu --say    # CPU (default)
ltts "Hello" -d cuda --say   # NVIDIA GPU
ltts "Hello" -d mps --say    # Apple Silicon

# Output formats
ltts "test" -o out.mp3       # MP3 (default)
ltts "test" -o out.wav       # WAV
ltts "test" -o out.ogg       # OGG
ltts "test" -o out.flac      # FLAC

# Language override
ltts "Bonjour" -l fr --say

Notes

• First run downloads models to ~/.cache/huggingface/ (~3GB for Qwen 1.7B, ~330MB for Kokoro)
• Audio playback (--say) runs at 24 kHz
• On Linux, ensure PulseAudio/PipeWire is running for audio playback

Development

uv sync
uv run ltts "hello world" --say
uv run ltts "hello world" -b kokoro -v af_heart --say
./scripts/release.sh