Supported Models

VoxServe supports multiple TTS and STS model families. The tables below summarize the current families, their shorthand codes, and a representative Hugging Face identifier.

Text-to-speech (TTS)

TTS model families

Family code	Example Hugging Face model ID	Description
chatterbox	ResembleAI/chatterbox (link)	TTS model developed by Resemble AI. Using 0.5B LLM with flow matching + HiFT vocoder. Supports audio input for voice cloning.
cosyvoice2	FunAudioLLM/CosyVoice2-0.5B (link)	TTS model developed by Alibaba. Using 0.5B LLM with flow matching + HiFT vocoder. Supports audio input for voice cloning.
csm	sesame/csm-1b (link)	TTS model developed by Sesame. Using 1B LLM and depth-wise model with Mimi detokenizer.
orpheus	canopylabs/orpheus-3b-0.1-ft (link)	TTS model developed by Canopy Labs. Using 3B LLM with SNAC detokenizer.
zonos	Zyphra/Zonos-v0.1-transformer (link)	TTS model developed by Zyphra. Using 1B LLM with DAC detokenizer.

Speech-to-speech (STS)

STS model families

Family code	Example Hugging Face model ID	Description
glm	zai-org/glm-4-voice-9b (link)	STS model developed by Z.ai. Using 9B LLM with flow matching + HiFT vocoder.
step	stepfun-ai/Step-Audio-2-mini (link)	STS model developed by StepFun. Using 8B LLM with flow matching + HiFT vocoder.

Notes

• The examples above are representative model IDs. You can use local paths or other compatible variants within each family.

• Some families support audio input (STS). Refer to the model card for input requirements.