--- title: "Alibaba's Fun-Realtime-TTS Beats Google to Claim the #1 Speech Throne" subtitle: "Alibaba's new TTS model edges past Google and Inworld on the blind-vote leaderboard with an Elo of 1,219, undercutting most rivals on price." url: "https://alphasignal.ai/news/alibaba-s-fun-realtime-tts-beats-google-to-claim-the-1-speech-throne" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "Artificial Analysis" company_url: "https://alphasignal.ai/company/artificialanlys" published: "2026-06-03T03:40:25.000Z" updated: "2026-06-04T23:09:55.074Z" type: "News" topics: - "Audio" - "Benchmarks" subtopics: - "Realtime Voice" - "Text To Speech" - "Voice Cloning" upvotes: 199 read_time: "4 min" source_url: "https://artificialanalysis.ai/text-to-speech/leaderboard/selected-voice" comments: 0 --- # Alibaba's Fun-Realtime-TTS Beats Google to Claim the #1 Speech Throne > Alibaba's new TTS model edges past Google and Inworld on the blind-vote leaderboard with an Elo of 1,219, undercutting most rivals on price. The top of the text-to-speech leaderboard has a new occupant from China. Alibaba's [Fun-Realtime-TTS](https://artificialanalysis.ai/text-to-speech/models/fun-realtime-tts) just climbed to the #1 spot on the Artificial Analysis Speech Arena, edging out Google's Gemini 3.1 Flash TTS and Inworld's Realtime TTS-2 Research Preview in blind listener comparisons. The Speech Arena is not a synthetic benchmark. Rankings come from blind user votes where listeners hear pairs of speech samples generated from the same text and pick which sounds more natural, then models are ranked using an Elo system. Higher scores indicate speech that listeners prefer more often. That makes the leaderboard a reasonable proxy for which models actually sound human to real ears. ## A photo finish at the top The headline number is tight. Fun-Realtime-TTS posted an Elo of 1,219 (+16/-16) across 962 arena appearances, placing it ahead of Gemini 3.1 Flash TTS at 1,214, Inworld Realtime TTS-2 Research Preview at 1,209, and Cartesia Sonic 3.5 at 1,203. Just 24 Elo points separate the top five models, which in Elo terms means listeners are essentially split down the middle on many of these head-to-head comparisons. This is also a meaningful jump for Alibaba. The lab's previous Fun-Realtime-TTS-Preview reached #7 on the leaderboard, making this Alibaba's first #1 model in the Artificial Analysis Speech Arena. The preview version had only cracked the global top five a few weeks earlier with an Elo of 1,190, so this is a sharp climb in a short window. ## What you actually get Fun-Realtime-TTS is the productized cousin of Alibaba's [Qwen3-TTS](https://github.com/QwenLM/Qwen3-TTS) work, and the feature list reflects the same engineering choices. The model supports real-time speech generation with voice cloning, voice design, multilingual output, and regional accents and dialects. A few details worth knowing if you are evaluating it for production: - **Voice cloning from short samples.** The underlying Qwen3-TTS architecture can clone a voice from just 3 seconds of audio, create new voices from text descriptions, and handle multiple Chinese dialects. It is trained on over 5 million hours of speech data and uses a dual-track architecture that keeps latency low, around 97 milliseconds for the first audio packet. - **Voice design vs cloning.** Voice cloning duplicates a specific voice from audio samples, while voice design creates a new voice from a text description such as a warm, low-pitched female voice. You use cloning when you have a target voice and design when creating one from scratch. - **Language coverage.** Ten mainstream languages including Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian, plus various dialects. - **Long-form stability.** The model can produce over 10 minutes of continuous audio without degrading quality or introducing artifacts. ## Pricing puts it in an awkward middle The commercial positioning is interesting. Fun-Realtime-TTS is priced at $27.59 per 1M characters, sitting between Gemini 3.1 Flash TTS at $18.3 and Inworld Realtime TTS 1.5 Max at $35, while remaining below Sonic 3.5 at $39. So Alibaba is not the cheapest option among the frontier models. Google is undercutting it by roughly a third while sitting only 5 Elo points behind. For latency-sensitive voice agents where quality is paramount, the price gap may not matter. For high-volume batch generation, it probably will. ## How to actually use it The model is delivered through Alibaba Cloud's Model Studio via the DashScope SDK. The realtime API uses a WebSocket connection for sending text and receiving streamed audio output, with a Python interface that looks roughly like this: ``` import os, dashscope from dashscope.audio.qwen_tts_realtime import * dashscope.api_key = os.environ.get('DASHSCOPE_API_KEY') config = QwenTtsRealtimeConfig.builder() \ .voice("Cherry") \ .mode("server_commit") \ .build() # stream text in, audio chunks arrive via callback ``` CosyVoice supports real-time speech synthesis and is available in the China (Beijing) and Singapore regions. Qwen-TTS supports both real-time and non-real-time synthesis, also available in both regions. The Singapore endpoint is the one to use if you are routing traffic outside mainland China. ## Why this matters beyond the leaderboard Voice is becoming the default interface for agentic applications, and the quality bar for sounding human, not robotic, is what makes or breaks deployments in customer support, in-car assistants, and accessibility tools. The fact that four different labs from three different countries are now within 24 Elo points of each other suggests the frontier of natural-sounding TTS is getting commoditized fast. For teams already shipping voice features on ElevenLabs, Cartesia, or Google, Fun-Realtime-TTS is now worth A/B testing, especially if multilingual or Chinese dialect coverage matters to your users. The [Speech Explorer](https://artificialanalysis.ai/text-to-speech/speech-explorer) has sample clips you can audition before writing any integration code. ## Sources - [X post by @artificialanlys](https://x.com/artificialanlys/status/2062016529848222073) - [Fun-Realtime-TTS: New Text to Speech model topping Artificial Analysis leaderboard](https://artificialanalysis.ai/articles/fun-realtime-tts-new-text-to-speech-model-topping-artificial-analysis-leaderboard) - [Text to Speech Leaderboard - Top AI Speech Models](https://artificialanalysis.ai/text-to-speech/leaderboard) - [Qwen3-TTS GitHub Repository](https://github.com/QwenLM/Qwen3-TTS) - [Qwen3 TTS Model Card on Replicate](https://replicate.com/qwen/qwen3-tts) - [Real-time speech synthesis - Alibaba Cloud Model Studio](https://www.alibabacloud.com/help/en/model-studio/realtime-tts-user-guide)