Sources
dots.tts: Apache-2.0 2B Open TTS Foundation Model Hits SOTA on Seed-TTS-Eval With 54ms Latency
The #1 trending HuggingFace Daily Paper (274 upvotes) introduces dots.tts, a 2B-parameter continuous autoregressive text-to-speech model achieving best average Seed-TTS-Eval performance (WER 0.94%/1.30% on zh/en) with strong voice cloning and emotional expressiveness. CFG-aware MeanFlow distillation yields first-packet latencies of 85ms (output streaming) and 54ms (dual-streaming). Training and inference code plus pretrained, post-trained, and distilled checkpoints are all released under Apache 2.0 — directly deployable for builders.
↳ Follow the thread