CuteTTS: Efficient and High-Quality Speech Synthesis via Autoregressive Modeling of Continuous Latents

A lightweight (~230M-parameter) continuous autoregressive TTS model that runs efficiently on GPUs, CPUs, and Apple silicon.
Ultra-low latency: ~40 ms to the first audio chunk and a throughput of ~9× real time on an NVIDIA RTX 4090.
Excellent speech quality and voice cloning performance.
Web demo, Python API, and CLI.
Multilingual support: English, Chinese, French, German, and Spanish.

Zero-shot voice-cloning performance

_Model	_Params.	_{LibriSpeech test-clean WER (%) ↓}	_{LibriSpeech test-clean SIM ↑}	_{Seed-TTS EN WER (%) ↓}	_{Seed-TTS EN SIM ↑}	_{Seed-TTS ZH WER (%) ↓}	_{Seed-TTS ZH SIM ↑}
_MOSS‑TTS	_8B	_1.98	_67.7	_1.84	_70.9	_1.37	_77.0
_Qwen3‑TTS	_1.7B	_2.35	_70.3	_1.66	_71.4	_0.91	_77.0
_{FireRedTTS‑2}	_1.5B	_4.32	_64.2	_1.95	_66.5	_1.14	_73.6
_{MOSS‑TTS‑Nano}	_0.1B	_4.10	_48.4	_4.62	_49.9	_3.13	_64.3
_F5‑TTS	_0.3B	_2.42	_66.0	_1.83	_67.0	_1.56	_76.0
_ZipVoice	_0.1B	_2.05	_67.4	_1.70	_69.7	_1.40	_75.1
_IndexTTS2	_1.5B	_2.47	_70.0	_2.22	_70.6	_1.02	_76.5
_CosyVoice 3	_0.5B	_1.99	_69.7	_2.02	_71.8	_1.16	_78.0
_VoxCPM2	_2B	_3.01	_74.0	_1.84	_75.3	_0.97	_79.5
_VibeVoice	_1.5B	_–	_–	_3.04	_68.9	_1.16	_74.4
_DiTAR	_0.6B	_2.39	_67.0	_1.69	_73.5	_1.02	_75.3
_{VibeVoice‑Realtime}	_0.5B	_2.00	_69.5	_2.05	_63.3	_–	_–
_Pocket TTS	_0.1B	_1.59	_49.1	_1.63	_50.7	_–	_–

_CuteTTS	_0.2B	_2.16	_78.9	_2.04	_76.5	_1.41	_77.8
_{CuteTTS‑distill}	_0.2B	_2.41	_76.8	_2.03	_74.2	_1.47	_75.6