FireRedTTS 2

FireRedTTS 2 tackles long-form conversational synthesis, generating podcast-style dialogues of up to 3 minutes with 4 distinct speakers while keeping speaker switches reliable and prosody consistent with conversational context. The system runs a dual-transformer over interleaved text and speech sequences on top of a 12.5 Hz streaming speech tokenizer, generating sentence by sentence, and reaches a 140 ms first-packet latency on an NVIDIA L20 in the team's tests, which suits live voice agents as well as offline production. It covers English, Chinese, Japanese, Korean, French, German, and Russian, and supports zero-shot voice cloning, including cross-lingual and code-switching scenarios, plus random timbre generation for building synthetic training data. Deployment means a Python 3.11 environment or Docker image with weights pulled from Hugging Face via Git LFS; bf16 inference brings VRAM needs from 14 GB down to about 9 GB. Code and weights from the Xiaohongshu FireRed team are Apache-2.0, though the README restricts the voice cloning feature to academic research, and a paper and demo page accompany the release.