Edge0 open-sources Audio8-TTS-Realistic-Air, a 0.6B-parameter speech model that speaks 23 languages and can copy a voice from one recording

Tue, Nov 3 15:09 UTC+8Updated Tue, Nov 3 18:32 UTC+8

Speech models are usually trained for clarity and voice match, missing the tone and rhythm of real people. Edge0 built Audio8-TTS-Realistic-Air on its Preview model, using a scoring model that compares its speech with human recordings to guide reinforcement learning. Languages grow from 11 to 23 and voice similarity from 0.682 to 0.704; code and weights are Apache 2.0. Quality varies by language.

Why it matters

Developers who want multilingual voiceovers or voice cloning on their own hardware can use this small open-source model directly, though quality varies by language.

Sources3

Hugging FaceMediaTue, Oct 27 14:55 UTC+8Edge0/Audio8-TTS-Realistic-AirView original · Tue, Oct 27 14:55 UTC+8
GitHubGitHubTue, Oct 27 18:04 UTC+8Edge0-AI/Audio8_TTSView original · Tue, Oct 27 18:04 UTC+8
arXivResearchSat, Oct 10 17:59 UTC+8Training a Compact, Strong, and Realistic Text-to-Speech ModelView original · Sat, Oct 10 17:59 UTC+8

Written by dalao.ai: the title, summary and why it matters are our own words. The original reports are the record; their publishers keep authorship.