Local Text-to-Speech Open Models: A 5-Minute Guide

Local TTS starts in 5 minutes, and your use case picks the model. The beginner default is Kokoro (82M) - Apache-licensed so commercial is fine, weights about 327MB, and it runs in real time even on a CPU with no GPU (a measured 5x on a 32-core CPU and 36x on a T4), in just a few lines of Python. But it cannot clone a voice (fixed voicepacks). If you need your own voice or multilingual cloning, you move to Chatterbox or XTTS - and here the license splits: Chatterbox is MIT (commercial OK) while XTTS v2 is CPML (non-commercial).

In one line: local TTS is chosen by use case - just-fast on CPU is Kokoro (Apache, no cloning), embedded is Piper, cloning plus commercial is Chatterbox (MIT), personal multilingual cloning is XTTS (CPML, non-commercial), and for Korean, Kokoro or Chatterbox are smoothest.

In plain terms: picking a TTS model is picking a ride. Kokoro is a bicycle (light, instant, anywhere), Piper is an e-scooter (ultralight, embedded), Chatterbox is your own car (it even clones a voice), and XTTS is a rental (multilingual but a no-commercial clause).

Which model fits your use case?#

==Just fast = Kokoro, embedded = Piper, cloning + commercial = Chatterbox, personal multilingual cloning = XTTS. RTF (real-time factor) below 1 means faster than real time (lower is better), and Kokoro is a measured ~0.03 on an A100. Korean has a trap: Kokoro supports Korean, and Chatterbox Multilingual V3 clones Korean across 23+ languages, but Piper has no official Korean voice== (community forks only). So for Korean, Kokoro (fixed voice) or Chatterbox (cloning) are the smoothest.

Below is the decision flow: two questions about your use (do you need cloning? is it commercial?) split the model.

How does the 5-minute install actually go?#

Mostly one pip line plus a few lines of Python. For Kokoro, install the system dependency espeak-ng, then pip install kokoro soundfile, and pass text, a voice (e.g. af_heart), and a speed to the pipeline to get 24kHz wav. If you dislike code, the Kokoro-FastAPI Docker image gives you an OpenAI-compatible API. Piper is one line, pip install piper-tts; Chatterbox is pip install chatterbox-tts and clones from a 5-second reference. But Chatterbox is Python 3.11 only, so it fails to install on newer Python - make a separate 3.11 venv (the most common beginner failure).

Kokoro's 5-minute path is short, as below: dependency, install, synthesize, and the first wav comes out in three steps.

Where do beginners get stuck?#

Three things: license, Korean, and the first run.

• License: XTTS is non-commercial (CPML), while Chatterbox, Kokoro, and Piper are commercial-OK. If you will sell it, check the license first.
• Korean: Piper has no official Korean, so use Kokoro or Chatterbox.
• First run and benchmarks: the first run is slow due to model download, so measure from the second, and "beats ElevenLabs"-style scores are usually vendor-run benchmarks, so measure on your own text.

How do you do it in 5 minutes?#

Start with the easiest path, and you can synthesize your first voice within five minutes.

• For just-fast, pip install Kokoro and synthesize one sentence to feel the RTF (CPU is fine).
• If you need your own voice, use Chatterbox (a Python 3.11 venv) with a 5-second reference. Outputs are watermarked.
• Do not clone anyone's voice without consent. For commercial use, check the license first (especially XTTS).

Reference links

• Kokoro (ultralight 82M TTS)
• Piper (embedded TTS)
• Chatterbox (cloning, MIT)
• Coqui TTS / XTTS v2
• Kokoro-FastAPI (no-code API)

Note: RTF, speed, and license details are public 2026 materials and measurements and vary by model, version, hardware, and language (published RTF is often an estimate, and "beats competitor" claims are usually self-run benchmarks). Measure exact speed and quality on your own text. Honor consent, license, and local regulation for cloning, and since models update often, this is reviewed quarterly.