Edge0 releases small voice-cloning model that speaks 23 languages for free

What you could do with it

• A developer could build a free voice-cloning tool that runs on affordable hardware since the model is small enough to operate on a standard CUDA GPU or even an Apple M2 CPU ¹.
• A content creator could generate speech in any of 23 languages by providing just one short audio clip and its transcript, using the model's zero-shot cloning feature ¹.
• A researcher could study, modify, and redistribute the model's code and weights freely because Edge0 released it under the Apache 2.0 open-weight license ¹.

What the model does

Audio8-TTS-Realistic is the second version in Edge0's small-model TTS line. It contains roughly 601 million parameters, excluding the audio codec. The company describes it as an open-weight model under the Apache 2.0 license, meaning developers can study, modify and redistribute the code.

The system takes a single reference recording and its transcript, then generates speech in the cloned voice. It supports 23 languages — a jump from the 11 languages in the earlier preview release. The latest releases additionally offer support for Turkish, Vietnamese, Swedish, Russian, Portuguese, Norwegian, Malay, Indonesian, Hindi, Finnish, Arabic, and Thai.

It can also create speech without any reference voice at all, functioning as a standard TTS system ¹.

How it is built

The architecture uses a dual-autoregressive design, meaning it predicts speech in two stages. The first stage runs on a model based on Qwen2.5 0.5B and produces text tokens plus the first layer of audio codes at roughly 21.5 frames per second. A second, shallower branch then fills in the remaining nine audio layers in parallel ¹.

Training happens in three phases. First comes standard supervised fine-tuning across many languages. The second phase uses a method called GRPO — a reinforcement learning approach — that rewards the model for low word error rates and good speaker similarity. The third phase adds a learned "realism critic" that scores outputs against real human recordings ¹.

Edge0 built a quality ladder using six open-source TTS systems ranked in a private listening study, with actual human speech at the top. An identity-erasure loss prevents the critic from simply rewarding mechanical fingerprints rather than genuine realism ¹.

A duration penalty blocks the system from gaming the critic by speaking unusually slowly. The final training stage used only Chinese and English prompts but still improved performance on lower-resource languages ¹.

How well it performs

On the Seed-TTS benchmark (English and Chinese), the model achieves 1.25% average error with a speaker similarity score of 0.704 ¹. On CV3, a nine-language test, it scores 4.86% average error — better than VoxCPM2 at 5.35% and Qwen3-TTS-1.7B at 7.73% .

On MiniMax, covering 24 languages, it reaches 5.89% error with 0.754 similarity ¹. On certain benchmarks, the bigger models stay ahead: Higgs TTS 3 achieves scores of 4.41% on CV3 and 2.74% on MiniMax, whereas Qwen3-TTS-1.7B records an average of 1.01% on Seed-TTS.

Weaknesses remain. Thai scores 48.57% word error on CV3, and Cantonese reaches 31.35% character error on MiniMax — far behind the best languages ¹.

Where to find it

The model is available on GitHub and Hugging Face. It requires PyTorch 2.5 or later and Transformers 4.57 or later. A CLI tool, batching support and fine-tuning scripts are included ¹.

For GPU serving, an SGLang adapter provides paged attention and dynamic batching with an OpenAI-compatible API endpoint. Edge0 notes that throughput and streaming speeds have not been benchmarked ¹. A free playground runs on Hugging Face ZeroGPU .

The model was smoke-tested on an NVIDIA H20 GPU in BF16 precision and also runs on Apple M2 CPUs using INT4 quantization, consuming roughly 1.1 to 1.2 gigabytes of memory ¹.

Key facts

• 601 million parameters with 23 supported languages
• CV3 error of 4.86%, beating Qwen3-TTS-1.7B and VoxCPM2
• Apache 2.0 open license
• Works on NVIDIA GPUs and Apple M2 CPUs
• Thai and Cantonese remain weak at nearly 49% and 31% error

Sources

1. Edge0 Releases Audio8-TTS-Realistic, a 0.6B Voice Cloning TTS Model — marktechpost.com