Kokoro

Kokoro-82M is an open-weight text-to-speech model by hexgrad with just 82 million parameters, derived from the StyleTTS 2 architecture. Despite its small size it delivers quality widely reported as comparable to much larger cloud TTS systems, while running fast enough for CPU inference. The v1.0 release ships 54 preset voices across 9 locales (American/British English, Japanese, Mandarin, Spanish, French, Hindi, Italian, Brazilian Portuguese) under Apache-2.0 weights, making it deployable in commercial products without API fees. The companion kokoro PyPI package and the hexgrad/kokoro inference library provide a simple Python pipeline; the ecosystem adds OpenAI-compatible servers (Kokoro-FastAPI), Rust, and CoreML ports.

Install the official inference library (Python 3.10–3.12):

```bash

pip install -q kokoro>=0.9.4 soundfile

# espeak-ng is required for grapheme-to-phoneme conversion:

apt-get -qq -y install espeak-ng

```

Basic usage:

```python

from kokoro import KPipeline

import soundfile as sf

pipeline = KPipeline(lang_code='a') # 'a' = American English

text = "Kokoro is an open-weight TTS model with 82 million parameters."

generator = pipeline(text, voice='af_heart')

for i, (gs, ps, audio) in enumerate(generator):

sf.write(f'{i}.wav', audio, 24000) # 24 kHz output

```

Weights (~82 MB) download automatically from Hugging Face on first run. Voice IDs encode locale + gender (e.g. af_heart = American female). For an OpenAI-compatible HTTP server, see the third-party remsky/Kokoro-FastAPI wrapper.