Alibaba Cloud CosyVoice 2

Alibaba Cloud CosyVoice 2 is a next-generation text-to-speech synthesis model released in December 2024, designed for speech agent builders who need ultra-natural voice output with zero-shot voice cloning from brief reference audio. It is a TTS-first cascade component: developers connect it to their own STT pipeline and LLM to form a complete speech-to-speech system, with Alibaba's DashScope SDK providing the WebSocket streaming transport. The model delivers sub-250ms streaming latency for real-time generation and is billed per input character at approximately $0.0000287 per character, translating to roughly $0.022 per minute at typical TTS output rates. CosyVoice 2 ships with over 100 built-in voices optimized for Mandarin Chinese, English, Japanese, and Korean, with Mandarin quality significantly exceeding comparable models from Western providers. Its zero-shot cross-lingual voice cloning capability allows production-quality cloned voices from a few seconds of audio without fine-tuning, making it ideal for personalized Mandarin-language assistants. The model is accessible via the DashScope Python and Java SDKs, with WebSocket streaming supported for real-time output. CosyVoice 2 is the leading choice for teams building Chinese-language voice assistants, Mandarin call center agents, or East Asian market products where natural-sounding Mandarin synthesis is non-negotiable. Its BYO-STT and BYO-LLM architecture gives developers full control over the input and reasoning layers while offloading voice synthesis to Alibaba's model. Teams outside the Alibaba Cloud ecosystem will need to manage cross-cloud data routing, and the 4-language ceiling limits its applicability for Western market products.