nkl-kokoro-tts-cli 0.1.9

nkl-kokoro-tts-cli 🗣️🤖

 

👉 Table of Contents

• 📍 Overview
• ✨ Features
• 📦 Installation
• 🔧 Development
• 🚀 Usage
  • Basic
  • Pipes (first-class)
  • mpv pipeline
  • No -o defaults
  • Formats
  • Voices & languages
  • Verbosity (Unix standard)
  • Long texts & splitting
  • Model & device
• 🐛 Troubleshooting
• 📤 Publish
• 🫂 Acknowledgments

📍 Overview

Minimal, pipeline-first CLI for Kokoro TTS (hexgrad/Kokoro-82M). Unix-philosophy: text in → wav/mp3 out, logs to stderr, stdout stays binary-clean.

✨ Features

• Pipeline-first — stdin/stdout everywhere. echo "hi" | kokoro-tts -o - | mpv --no-video - works, no temp file. Logs ([INFO]/[WARN]/[ERROR]) go to stderr only.
• Always local model — default ~/.cache/huggingface/hub (or $HF_HUB_CACHE/$KOKORO_MODEL_DIR). First run auto-downloads hexgrad/Kokoro-82M (~340M) via huggingface_hub, then fully offline. kokoro-tts download --model-dir ./kokoro-model for a plain dir.
• Unix verbosity — default WARN (only warnings/errors). -v → INFO, -vv → DEBUG, -q → ERROR only, --log-level overrides. Quiet by default for | mpv.
• Perf — in-process pipeline + voice tensor cache (cold 5.3s → warm 0.42s 12.5×), lazy torch/kokoro import so --help is instant, CUDA ~10× for long texts.
• Audio — wav default (no deps) via soundfile 24kHz, mp3 via ffmpeg (-f mp3), mpv pipe friendly.
• Voices & langs — am_echo default, 50+ known (af_, bf_, ef_, jf_, zf_…), auto lang from voice prefix (a US, b GB, e es, f fr, h hi, i it, p pt-br, j ja, z zh), custom .pt voice.
• Linux standards — uv + hatchling src/ layout, pyproject.toml, requires-python >=3.10, kokoro-tts entry point, 88 mocked tests, SIGPIPE safe.

📦 Installation

System deps (recommended):

# Debian/Ubuntu
sudo apt install espeak-ng ffmpeg mpv

# Arch
sudo pacman -S espeak-ng ffmpeg mpv

espeak-ng improves G2P fallback, ffmpeg needed for mp3, mpv for pipe playback.

User install:

# uv tool (isolated, recommended)
uv tool install nkl-kokoro-tts-cli
kokoro-tts --help

# pip
pip install nkl-kokoro-tts-cli

# first run auto-downloads model to ~/.cache/huggingface/hub (~340M)
kokoro-tts "Hello world" -o hello.wav

Pre-download (optional, for offline):

kokoro-tts download                          # → ~/.cache/huggingface/hub
kokoro-tts download --model-dir ./kokoro-model  # → plain ./kokoro-model
KOKORO_MODEL_DIR=./kokoro-model kokoro-tts "Hi" -o out.wav

🔧 Development

git clone https://github.com/nkl/kokoro-tts-cli
cd kokoro-tts-cli

# env + deps (includes en_core_web_sm 3.8.0 via direct ref)
uv sync --extra dev
uv run kokoro-tts --help

# tests (70+ mocked, no GPU/HF needed)
uv run pytest -q          # 88 passed
uv run pytest -v
uv run pytest -k test_model_local

# lint / build / publish
uv run ruff check src tests
uv build
uv run twine check dist/*
uv tool install --force dist/nkl_kokoro_tts_cli-*.whl

Project layout: src/nkl_kokoro_tts_cli/{cli,tts,logger}.py (cli = pipe/arg parsing, tts = Kokoro wrapper + pipeline/voice cache, logger = leveled stderr), tests/{test_cli,test_tts,test_logger,test_model_local}.

🚀 Usage

kokoro-tts --help
kokoro-tts --list-voices
kokoro-tts --version

Basic

kokoro-tts "Hello world, this is Kokoro." -o hello.wav
kokoro-tts "Hello" --voice am_echo --speed 1.0 -o hello.wav
kokoro-tts "Hello" --voice af_bella --lang a -o out.wav  # lang auto from voice

Pipes (first-class)

echo "Hello from the pipeline" | kokoro-tts -o out.wav
cat book.txt | kokoro-tts --format mp3 -o book.mp3
kokoro-tts --input-file chapter.txt -o chapter.wav --speed 0.95
from_clipboard | kokoro-tts -o - | mpv --no-video -  # from_clipboard = xsel -ob

kokoro-tts writes wav to stdout when -o - or when stdout is piped and no -o given.

mpv pipeline

# wav (no ffmpeg)
echo "Hello, this is a test" | kokoro-tts --voice am_echo -o - | mpv --no-video -

# mp3 (needs ffmpeg)
echo "Hello" | kokoro-tts -o - --format mp3 | mpv -

# with speed / volume, silent logs
cat transcript.txt | kokoro-tts --voice af_bella --speed 0.95 -o - --log-level ERROR | mpv --no-video --volume=80 --really-quiet -

Tip: mpv --no-video - reads stdin. Our logs stay on stderr.

No -o defaults

• stdout is a pipe → wav to stdout (| mpv - works without -o -)
• stdout is a TTY → ./output.wav + [WARN]

Formats

kokoro-tts "Hello" -o out.wav               # wav (default)
kokoro-tts "Hello" -o out.mp3 --format mp3  # mp3 (ffmpeg)
echo "hi" | kokoro-tts -o - --format mp3 | mpv -

Voices & languages

kokoro-tts --list-voices                    # 50+ af_/am_/bf_/bm_/ef_/jf_/zf_...
kokoro-tts "Bonjour le monde" --voice ff_siwis -o fr.wav
kokoro-tts "Hola mundo" --voice ef_dora --lang e -o es.wav
kokoro-tts "Hello" --voice /path/to/custom.pt -o out.wav
kokoro-tts "Hi" --voice af_bella,af_nicole -o averaged.wav  # averaged

--lang auto-inferred from voice prefix; default am_echo → a (US).

Verbosity (Unix standard)

kokoro-tts "Hello" -o out.wav               # default WARN (only WARN/ERROR)
kokoro-tts "Hello" -o out.wav -v            # INFO + WARN + ERROR
kokoro-tts "Hello" -o out.wav -vv           # DEBUG (adds chunk/ps debug)
kokoro-tts "Hello" -o out.wav -q            # ERROR only
kokoro-tts "Hello" -o out.wav --log-level DEBUG

Long texts & splitting

Kokoro chunks at 510 phonemes. Override:

kokoro-tts --input-file long.txt --split-pattern '\n+' -o out.wav

Model & device

# model (always local)
kokoro-tts download --model-dir ./kokoro-model  # plain dir
KOKORO_MODEL_DIR=./kokoro-model kokoro-tts "Hi" -o out.wav
# default HF cache: ~/.cache/huggingface/hub/models--hexgrad--Kokoro-82M

# device (auto picks CUDA if available)
kokoro-tts "Hello" -o out.wav --device auto -v  # [INFO] pipeline ready device='cuda:0' (hf-cache)
kokoro-tts "Hello" -o out.wav --device cpu
kokoro-tts "Hello" -o out.wav --device cuda

Long 2600 chars: CPU 40s vs CUDA 4s.

🐛 Troubleshooting

• ffmpeg not found for mp3 → sudo apt install ffmpeg
• EspeakFallback not Enabled → sudo apt install espeak-ng
• Can't find model 'en_core_web_sm' → fixed in 0.1.5+ via direct en_core_web_sm 3.8.0 dep; uv tool update nkl-kokoro-tts-cli
• unauthenticated requests to HF Hub → suppressed in 0.1.5; set export HF_TOKEN=hf_xxx to raise rate limits
• CUDA requested but not available → --device cpu

📤 Publish

uv build
uv run twine check dist/*
uv publish

Or with just:

just build   # uv build (runs tests first)
just publish # uv build + twine check + uv publish

pyproject.toml uses hatchling, allow-direct-references for en_core_web_sm, entry kokoro-tts = nkl_kokoro_tts_cli.cli:main.

🫂 Acknowledgments

• SUP3RMASS1VE/Ultimate-TTS-Studio-SUP3R-Edition, where I discovered the handy kokoro-tts.
• hexgrad/kokoro, for the interface built on top of Kokoro-82M.

 

⭐ Hit the Star Button if You Found This Useful ⭐