OpenMOSS Team releases MOSS-TTS, an 8B Apache 2.0 voice cloning and long-form TTS

OpenMOSS Team released MOSS-TTS, a family led by an 8B flagship (MossTTSDelay, recommended for production) and a 1.7B MossTTSLocal (recommended for research). The card lists 20 languages but its table shows 19. Features include zero-shot voice cloning from reference audio, long-form generation described as up to one hour in a single run, token-level duration control (about 12.5 audio tokens per second), Pinyin and IPA pronunciation input, and code-switching within one utterance. Siblings include MOSS-TTSD for dialogue, MOSS-VoiceGenerator for voice design, MOSS-SoundEffect and MOSS-TTS-Realtime, which reports 180 ms time to first byte.

Self-reported Seed-TTS-eval: MossTTSDelay scores 1.84 English word error rate and 1.37 Chinese character error rate. The card calls these state of the art, but its own table shows Qwen3-TTS at 1.5 on English and GLM-TTS-RL at 0.89 on Chinese, both better.

The licence is Apache-2.0. Practical notes: it needs trust_remote_code, the README recommends Transformers 5.0.0, and only one reference audio is supported. The card gives no release date; the date used is the Hugging Face repo creation date, and the linked report is listed as published 18 March.

Useful family. The card's own table keeps the claims honest.

Sources

• [OpenMOSS Team model card on Hugging Face](https://huggingface.co/OpenMOSS Team-Team/MOSS-TTS)
• [Hugging Face repo metadata (creation date)](https://huggingface.co/api/models/OpenMOSS Team-Team/MOSS-TTS)

Editor's first pass (AI-assisted)

• Real 4: Apache 2.0 weights, though the card needs trust_remote_code and pinned torch.
• Useful 4: Long-form output, duration control and cloning cover narration and dubbing work.
• Notable 3.5: A broad family with dialogue, sound effects and realtime siblings.
• Novel 3.5: Token-level duration control is uncommon, the rest is familiar.