Flare-TTS 28M debuts from-scratch TTS model

A tiny, from-scratch TTS release like this is more interesting as an engineering proof point than a production voice engine. It shows how far a solo builder can push speech synthesis on modest hardware, even if the output still has a rough edge.

• –28M parameters makes this firmly a small-model story, not a state-of-the-art voice clone story
• –Training on one A6000 in roughly 24 hours makes the setup accessible to independent researchers and hobbyists
• –LJSpeech gives it a narrow English-only scope, so generalization and multilingual quality are not the point here
• –The release is useful for anyone studying compact TTS architectures, low-budget training, or local speech tooling
• –The “robotish” quality note is a realistic signal: this is a base release worth watching, not a polished consumer voice product