How to Run Kokoro-82M Locally: the Fastest Open-Weight TTS Model

Kokoro-82M is an 82 million parameter open-weight text-to-speech model, fully Apache 2.0 on both code and weights, with 54 voices across 8 languages. At fp16 the weights are roughly 164 MB, small enough that it runs comfortably on a CPU or almost any GPU. Install it with one pip command and generate speech in a few lines of Python.

Why Kokoro-82M gets so much traffic

Kokoro-82M pulls roughly 11.6 million downloads a month on Hugging Face and carries 6.82 thousand likes, more usage than any other open text-to-speech model. The draw is the size-to-quality ratio: a StyleTTS 2 architecture with an ISTFTNet decoder, no diffusion step, trained on a few hundred hours of permissive audio, outputting 24 kHz PCM. Training reportedly cost about 1,000 A100 GPU hours, roughly 1,000 dollars at market rates, a fraction of what most speech models spend. Community API providers price it under one dollar per million characters, which is what happens when a model is small enough to run cheaply almost anywhere.

Kokoro-82M vs Breeze TTS 2

The two models solve different problems. Breeze TTS 2 clones and directs a specific voice at 3 billion parameters, needs a CUDA GPU with at least 12 GB of VRAM, and licenses its weights for non-commercial use only. Kokoro-82M is a thousand times smaller, ships a fixed set of 54 stock voices instead of cloning, runs on a CPU if you have to, and its Apache 2.0 license has no commercial restriction at all. Reach for Breeze TTS 2 when you need to match a specific voice; reach for Kokoro-82M when you need cheap, fast, unrestricted speech and any clear voice will do.

Install Kokoro-82M

Kokoro ships as a pip package. It leans on espeak-ng for phonemization, so install that too before your first run.

Generate your first line of speech

The KPipeline class takes a language code and hands back a generator. Each step yields the grapheme string, the phoneme string, and the audio array for one chunk of text.

from kokoro import KPipeline
import soundfile as sf

pipeline = KPipeline(lang_code="a")
text = "Your text here"
generator = pipeline(text, voice="af_heart")

for i, (graphemes, phonemes, audio) in enumerate(generator):
    sf.write(f"output_{i}.wav", audio, 24000)

Languages and voices

The lang_code argument picks the phonemizer and voice set. 54 voices are split across the 8 supported languages; af_heart, used above, is an American English voice.

Can you use it commercially?

Yes, without restriction. Both the weights and the inference code are Apache 2.0, so Kokoro-82M can run in a commercial product, a hosted API, or a personal project on the same terms.

Try the voices first in the free Kokoro-TTS Space on Hugging Face before you install anything.

FAQ

Does Kokoro-82M need a GPU?

No official benchmark numbers are published, but at 82 million parameters the fp16 weights are only about 164 MB, small enough to run on a CPU or any GPU with a couple of gigabytes free.

What voice should I start with?

af_heart is the voice used in Kokoro's own quick-start example and a reasonable default for American English. Check VOICES.md for the other 53 options across all 8 languages.

Can I use Kokoro-82M in a commercial product?

Yes. Both the weights and the code are Apache 2.0 with no non-commercial clause, unlike some other open text-to-speech releases.