# Free local AI voice for your videos

> How to generate a narration voice in English or Italian on your own computer with Kokoro, an open model: no online service and no subscription.

Status: testato · Updated: 2026-11-20

**TL;DR:** Kokoro-82M is a small open speech model: it runs on a laptop CPU, has English and Italian voices and sends nothing online.

Videos with puppets need a voice. Online voices cost money, send your text to an outside service and sometimes change license. A model that runs on your computer solves all of that: the text stays yours and there is no subscription.

## Install
On Linux you need `espeak-ng` for phonetics (`sudo apt install espeak-ng`). Then, in a Python environment:

```bash
python -m venv .venv && source .venv/bin/activate
pip install kokoro soundfile numpy
```

The first run downloads the model (a few hundred MB). After that it also works offline.

## One voice per scene
One sentence per scene, and the script tells you how long each one lasts. Those durations are what you use to match scenes to the voice (see [syncing scenes and voice](/en/guide/syncing-scenes-and-voice)).

```python
import numpy as np, soundfile as sf
from kokoro import KPipeline

SR = 24000
pipe = KPipeline(lang_code="i")          # "a" = American English, "i" = Italian
scene_texts = ["You stole a car and the police are after you? Relax.", "Swap the car and change clothes."]

audio, t = [], 0.0
for i, text in enumerate(scene_texts):
    parts = [np.asarray(a, dtype=np.float32) for _, _, a in pipe(text, voice="af_heart")]
    wav = np.concatenate(parts)
    pause = np.zeros(int(0.4 * SR), dtype=np.float32)   # a little breathing room between scenes
    dur = (len(wav) + len(pause)) / SR
    print(f"scene {i+1}: from {t:.2f} s to {t+dur:.2f} s")
    audio += [wav, pause]
    t += dur
sf.write("voce.wav", np.concatenate(audio), SR)
```

The voices I tried are `af_heart` and `am_michael` for American English, and `if_sara` (female) and `im_nicola` (male) for Italian. Change `lang_code` together with the voice.

## What to expect
The voice is clear and fast, but flat: no emotion, and you cannot ask it to "say it with more energy". For an explainer video that is fine; for storytelling you need slower models that handle expressiveness. Write short sentences with proper punctuation, because the model reads the way the text reads: an extra comma is an extra pause.

Listen to everything before publishing. Numbers, acronyms and foreign names (Leonida, Vice City) sometimes come out wrong: if that happens, respell them the way they sound.

To use the voice in a video see [puppet videos without any gameplay](/en/guide/puppet-videos-without-gameplay).

*Tested on 20 November 2026 on a laptop without a graphics card: two Italian sentences, about 15 seconds.*

## FAQ

### Do I need a graphics card for an AI voice?
Not for Kokoro: it is small (82 million parameters) and runs on the CPU. In one test, two Italian sentences took about 15 seconds, model loading included.

### Can I use these voices in monetized videos?
Check the license of the model and of the voices you download. Kokoro is published under a permissive license, but verify on the model page before monetizing.

## Sources
- [Kokoro-82M (Hugging Face)](https://huggingface.co/hexgrad/Kokoro-82M)
