THE ESSENTIALS IN 30 SECONDS

  • Audio8-TTS-Realistic has about 601 million parameters (excluding its codec) and is released under Apache 2.0 for both code and weights.
  • It supports 23 languages, 12 more than the earlier Preview release, and clones a voice zero-shot from a reference recording and its exact transcript.
  • The model card asks users to get permission before cloning a voice and to disclose synthetic audio, but lists no technical safeguards or formal safety evaluation.

Edge0, a small company that builds models meant to run locally on phones, laptops, wearables and robots, has published Audio8-TTS-Realistic, a text-to-speech model that can imitate a speaker from a single reference recording. The release, which drew online coverage at the end of October, is now the main model in the company's Audio8 TTS repository, replacing an earlier Preview version.

Size, licence and languages

According to its Hugging Face model card, the model has 601 million parameters, not counting the audio codec shipped alongside it. Code and weights are both under the Apache 2.0 licence, a permissive licence that requires attribution, listed in a NOTICE file. It is a fine-tune of Audio8-TTS-Preview-0.6b and uses a two-stage autoregressive design that the card says was inspired by Fish Audio S2 Pro: a 24-layer transformer predicts one semantic token per audio frame, and a smaller four-layer one fills in the codec details. Output is 44.1 kHz audio.

Language coverage grows from 11 to 23. The Preview set (including English, Chinese, Cantonese, Japanese, Korean, French, Spanish and German) is joined by Arabic, Hindi, Indonesian, Malay, Portuguese, Russian, Thai, Turkish, Vietnamese and three Nordic languages. Edge0 warns that quality varies from one language to another. Unlike the smaller 0.1B checkpoint, which runs on a CPU via ONNX, the new model is recommended for CUDA GPUs and has no ONNX export.

Benchmarks and safeguards

The company reports a speaker similarity score of 0.704 on the English and Chinese Seed-TTS test, up from 0.682 for Preview, with a nearly unchanged error rate of 1.25%. Its own comparison table shows other systems doing better on some tests, such as Qwen3-TTS-1.7B on Seed-TTS errors and Higgs Audio v3 on a 24-language MiniMax set. Edge0 stresses that these figures measure intelligibility and similarity, not how realistic the voices sound, and that cross-project comparisons use different pipelines.

Cloning requires a reference clip plus a transcript that must match it exactly. On misuse, the card's responsible-use section asks people to obtain permission before cloning someone's voice and to disclose synthetic audio where appropriate. Beyond that guidance, neither the model card nor the GitHub repository describes watermarking, consent checks or a formal safety evaluation, which matters for a freely licensed model capable of imitating real voices.