dots.tts Technical Report

Researchers have developed dots.tts, a 2-billion parameter text-to-speech model that achieves state-of-the-art performance through innovations in continuous speech modeling, full-history conditioning, and self-corrective training. The model demonstrates exceptional multilingual capabilities and enables low-latency speech generation, with code and weights released open-source under Apache 2.0 license.

The dots.tts release represents a significant advancement in open-source text-to-speech technology, addressing key limitations in current continuous autoregressive TTS models. The technical innovations—particularly the AudioVAE with multiple objectives and reward-free self-corrective post-training—demonstrate a systematic approach to improving both speech quality and generation stability. These methods create a semantically structured latent space that balances naturalness with predictability, a challenge that has plagued prior approaches.

This development builds on the broader AI trend toward larger, more capable foundation models combined with efficient deployment strategies. The field has progressively moved from discrete tokenized approaches toward continuous representations, recognizing that speech's inherent analog nature is better captured in continuous spaces. dots.tts advances this frontier while maintaining practical inference speed—achieving 85ms and 54ms latencies in different streaming modes—making it viable for real-world applications.

The open-source release carries significant implications for AI democratization and developer accessibility. By providing training code, multiple checkpoint versions, and permissive licensing, the developers enable downstream research and commercial applications that would otherwise require proprietary solutions. This accelerates innovation in voice cloning, multilingual speech synthesis, and emotional expressiveness—capabilities increasingly demanded across communications, content creation, and accessibility applications.

The model's strong multilingual performance (evidenced by competitive results across Chinese, English, and challenging variants) suggests robust cross-linguistic learning mechanisms. Moving forward, attention will focus on whether these architectural innovations become standard practice in TTS, how efficiently teams can fine-tune for specialized domains, and whether similar approaches translate to other audio generation tasks.

• →dots.tts achieves state-of-the-art multilingual TTS performance with error rates of 0.94% on Chinese and 1.30% on English
• →Low-latency inference design enables practical deployment with first-packet latencies of 54-85ms across streaming modes
• →Full-history conditioning and reward-free self-corrective training represent methodological advances applicable to other autoregressive models
• →Open-source Apache 2.0 release with complete code and checkpoints significantly lowers barriers to TTS research and commercialization
• →Strong voice cloning and emotional expressiveness capabilities position the model for creative and accessibility applications