Audio8-TTS-Realistic: a small open-weights voice cloning model in 23 languages
Audio8-TTS-Realistic is a free, open-weights text-to-speech model from Edge0 with 0.6 billion parameters that clones a voice from one short reference recording. It speaks 23 languages, including Polish, and the Apache 2.0 licence allows commercial use.
What it does
Give it a clean reference recording plus its exact transcript and it reads new text in that voice, or it can speak without a reference. Compared with the earlier Preview model, Edge0 says it sounds more natural and adds 12 languages (Arabic, Finnish, Hindi, Indonesian, Malay, Norwegian, Portuguese, Russian, Swedish, Thai, Turkish and Vietnamese) to the 11 it had. The download includes a 44.1 kHz audio codec, so no extra checkpoint is needed, and there is an SGLang Omni adapter with an OpenAI-compatible speech API for serving.
Worth knowing
It runs through Hugging Face Transformers with custom code (trust_remote_code), and a CUDA GPU is recommended; Edge0 smoke-tested it with PyTorch 2.8 on an NVIDIA H20 in BF16. The reference transcript must match the recording exactly. Edge0 notes that quality varies by language and publishes per-language scores. You can try it free in a Hugging Face Space, which runs on a shared ZeroGPU queue. Edge0's team says it is #1 on the Hugging Face TTS Arena, a blind listener vote.