Chinese TTS — side by side
Lesson 1 vocabulary, generated once on the VPS (4 vCPU, CPU-only). Same source text for both engines.
ZipVoice-Distill INT8 · female ref clip
Kokoro-82M v1.1 zh INT8 · zf_001
Full sentence
《汉语教程》lesson 1: 张东打电话找田芳，田芳的妈妈接电话说田芳不在家。她四点多就去同学家了。
ZipVoice
Kokoro (zf_001)
Measurements on this VPS (4 vCPU, int8, 4 threads)
Kokoro: 7.6 s of speech in 14.3 s → RTF 1.88. ZipVoice: 8.5 s in 19.5 s including ~5 s model load → RTF ≈ 1.7.
Both are slower than real time, so audio must be generated once and cached (per-word clips: 0.1–0.9 s each).