Opensourcing TADA: Fast, Reliable Speech Generation Through Text-Acoustic Synchronization

The future of voice AI is significantly advanced by TADA (Text-Acoustic Dual Alignment), a novel tokenization schema developed by Hume AI that synchronizes text and speech in a one-to-one alignment, addressing the mismatch in text and audio representation in language models. This innovation allows TADA to deliver the fastest LLM-based TTS system with competitive voice quality and virtually zero content hallucinations, suitable for on-device deployment. By representing audio with continuous acoustic vectors aligned to text tokens, TADA enhances speed and reduces computational effort, with evaluations showing it generates speech more than five times faster than similar systems and achieves high reliability with zero hallucinations. The model excels in context efficiency, supporting long-form and conversational speech while maintaining production reliability, making it ideal for applications in sensitive environments like healthcare and finance. Despite some limitations in long-form degradation and a modality gap during text generation alongside speech, TADA's open-source availability promises potential for further development and application expansion, with ongoing efforts to broaden language coverage and enhance model capabilities.