What for: an open TTS engine with voice cloning and 17+ languages; the weights live on even though the company behind it shut down.

What it is and who it's for

Coqui XTTS-v2 is a mature open TTS engine with voice cloning and multilingual output (Russian included). The company shut down, but the code and weights stayed open-source and are still widely used. For a developer who wants a time-tested local TTS with cloning and many languages in one toolkit. What sets it apart: the breadth of languages, and the fact that it's a "forever" open model with no vendor attached.

Key features

Voice cloning from a short sample (cross-language).
Multilingual: 17+ languages, Russian among them.
Open weights and the TTS toolkit — local, free, offline.
Low-latency streaming on a GPU.
Fine-tuning on your own data.
A rich Python API and CLI.

Get started in 5 minutes

Install the toolkit:

pip install coqui-tts

Synthesize with cloning and a chosen language:

from TTS.api import TTS
tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to("cuda")
tts.tts_to_file(text="Привет, мир", speaker_wav="sample.wav", language="ru", file_path="out.wav")

Swap in your own sample.wav and the ru language code.

When to take it, and when not

✅ Take it if: you need a multilingual local TTS with cloning and solid Russian.
✅ Take it if: you want a "forever" open model that doesn't depend on a living vendor.
❌ Skip it → "Chatterbox" is better if you want the freshest models and an actively developed project.
❌ Skip it → "Kokoro TTS" is better if you don't need cloning and care about being lightweight and cheap on resources.

The honest price

Free (open-source). Heads-up: the original XTTS-v2 shipped under a special license (Coqui Public Model) — check the terms for commercial use in the repo or forks. Your cost is the GPU. Exact requirements depend on your hardware.

Gotchas

The project has no official support (the company is gone) — bug fixes rest on the community and forks.
The model license is non-standard — for a commercial product, verify it first.
Clone quality depends on how clean the sample is; noise hurts.
Real-time needs a GPU; on CPU it is noticeably slower.

🤖 Prompt accelerator

"I'm taking Coqui XTTS-v2 for local multilingual voiceover with cloning (language: Russian). Help me install the toolkit, prepare a clean speaker_wav, write a tts_to_file script with language="ru", check the license terms for commercial use, and estimate GPU requirements. Tell me how to fine-tune the model on my own data if the clone quality isn't good enough."