OpenMOSS /MOSS-TTSD

About MOSS-TTSD

MOSS-TTSD is an open text-to-speech model built for dialogue, not one-voice narration. Give it a short clip of each speaker, then a script tagged [S1] to [S5]. It keeps each voice steady for up to 60 minutes of audio. The documented setup needs a GPU with flash-attention.

Best for: Voice directors who dub multi-speaker scripts into podcasts, audiobooks or drama dialogue.

• Clones 1 to 5 voices from short reference clips, no training needed
• Keeps each speaker's voice steady for up to 60 minutes in one run
• Speaks 20 languages, including Chinese, English and Japanese
• Runs batch jobs from a JSONL file of tagged dialogue scripts
• Serves the model through SGLang for faster generation
• Ships fine-tuning code for full-parameter and LoRA training (v0.5)

1. 1
   Create the environment: the command below

   conda create -n moss_ttsd python=3.12 -y && conda activate moss_ttsd

   Copy
2. 2
   Install dependencies: the command below then pip install flash-attn

   pip install -r requirements.txt

   Copy
3. 3
   Write a JSONL file with [S1]–[S5] speaker tags and matching prompt_audio_speakerN / prompt_text_speakerN reference pairs
4. 4
   Run batch inference: the command below

   python inference.py --model_path OpenMOSS-Team/MOSS-TTSD-v1.0 --codec_model_path OpenMOSS-Team/MOSS-Audio-Tokenizer --input_jsonl /path/to/input.jsonl --save_dir outputs --mode voice_clone_and_continuation --batch_size 1 --text_normalize

   Copy
5. 5
   Optional: serve the fused model with SGLang and send requests via the command below

   python scripts/request_sglang_generation.py

   Copy