Fun-Realtime-TTS (also known as Fun-Realtime-TTS-Preview) is a high-performance text-to-speech model developed by Alibaba's Tongyi Lab as part of the FunAudioLLM framework. It is optimized for ultra-low latency applications, such as real-time voice agents and interactive chatbots, achieving time-to-first-audio delays of approximately 97 to 150 milliseconds. In mid-2026, the model gained industry recognition by reaching the top position on the Artificial Analysis Speech Arena leaderboard, a benchmark determined by blind human preference.
The model's architecture utilizes a Dual-Track modeling approach that combines a large language model (LLM) backbone with a chunk-aware flow matching (FM) mechanism. This design allows it to handle both text-in and audio-out streaming concurrently, delivering the first audio packet after processing as little as a single character. It incorporates a proprietary 12Hz multi-codebook speech encoder that achieves high-fidelity voice compression while preserving paralinguistic nuances like emotion, tone, and prosody.
Key capabilities of Fun-Realtime-TTS include zero-shot voice cloning, which can replicate a speaker's identity from a 3-second audio sample, and voice design, allowing users to generate new voices using natural language descriptions. The model supports more than 30 languages and is noted for its extensive coverage of regional accents and over 20 Chinese dialects, such as Cantonese and Sichuanese. It can adaptively adjust rhythm and emotional expression based on the semantic context of the input text.
Official recommendations for voice design suggest using detailed, descriptive prompts (e.g., "a clear-toned young female voice with gentle intonation") rather than subjective terms or celebrity names to achieve consistent results. The model family is released in multiple sizes, including a 1.7B parameter version for peak performance and a 0.6B version optimized for edge efficiency.