--- title: "China's AI music race: three homegrown models take on Suno in Mandarin" date: 2026-10-03 category: Generative Media site: NeuroAI canonical: https://neuroai.site/a/na-aigc-ai-music-china language: en --- # China's AI music race: three homegrown models take on Suno in Mandarin > As Suno made AI songwriting a global talking point, Chinese companies built music large models tuned for Mandarin, dialects, and short-video creators. A musician hums a melody into her phone and gets back a finished backing track in her own dialect. A short-video creator types a mood and a length, and a matching song drops into the timeline. A small brand skips the licensing fees and generates a jingle that fits its ad exactly. ## The Suno shock, and the local answer In March 2024, the U.S. tool Suno released V3 and was quickly called the "ChatGPT moment" for music. By the time Suno V4 arrived, it could produce songs up to four minutes long. The release lit a fire under Chinese labs, which had watched the text and image races closely and did not want to be late to audio. What followed was not one product but a cluster. Chinese state media later grouped the field into three kinds of players: big-tech apps, independent model companies, and vertical specialists. The names that keep coming up are 海绵音乐 (Haimian Music) from ByteDance, 天工SkyMusic (SkyMusic) from Kunlun Tech (昆仑万维), and 天谱乐 (TianPuLe) from Tuanyi Tech (趣丸科技). ## 海绵音乐 (Haimian): the ByteDance path 海绵音乐 is ByteDance's consumer music tool, tied to the Doubao (豆包) assistant. The pitch is simple: type a line of inspiration or upload an image, and the model returns a personalized song. Reporting describes its strength as Chinese-language clarity — Mandarin pronunciation comes through more cleanly than on Suno, and the singing flow fits local tastes better. Output lands around one minute per song, and the basic tier is free, which matters for the casual creator who just wants a birthday clip or a Douyin (抖音) background track. ## 天工SkyMusic (SkyMusic): the architecture play Kunlun Tech (昆仑万维) opened 天工SkyMusic to a free invite test on April 2, 2024, and to the public on April 17, 2024, alongside its 天工3.0 (Tiangong 3.0) large model (大模型). 天工3.0 carried about 400 billion parameters and was promoted as the largest open-source mixture-of-experts model then available. Technically, SkyMusic is interesting. It uses a music-audio architecture described as Sora-like: a large-scale Transformer handles composition by learning the context of music patches, while a Diffusion Transformer handles the singing, turning those patches back into high-quality audio. The headline spec is the ability to generate 80-second songs at a 44,100 Hz sample rate in two-channel stereo. It also learns singing techniques — vibrato, opera, call-and-response duets — and supports dialect songs, which is where it pulls ahead of foreign tools for Chinese users. Kunlun did not stop at one model. In August 2024 it launched Melodio, a streaming-style music platform that generates continuous custom tracks from a mood or scenario, and Mureka, a commercial music-creation platform. ## 天谱乐 (TianPuLe): the multimodal specialist Tuanyi Tech's 天谱乐 arrived in July 2024 as a multimodal scoring model. Where the others lean on text or audio prompts, 天谱乐 can take an image or a video and produce matching music, pulling in picture-understanding and video-understanding algorithms alongside melody generation. For editors who already have visuals and need a score, that direction is more useful than a blank text box. The reach is not trivial. Reporting cited company figures of roughly 46 million registered users across Tuanyi's 唱鸭 (ChangYa) app and the 天谱乐 site, with close to 10 million AI songs created. Those are vendor-reported numbers, so treat them as directional. ## Why China's music models have an edge at home The global tools are strong, but Mandarin is a hard target. Tonal languages punish sloppy pronunciation, and local users notice immediately. The domestic models were trained and tuned for exactly this: clearer Chinese diction, support for regional dialects, and 国风 (guofeng, traditional-style) melodies that resonate locally. For the massive short-video (短视频) economy, a tool that sings in fluent Mandarin with the right emotional tone is more than a novelty — it is a content-production input. ## The copyright shadow No account of AI music is complete without the legal cloud. In June 2024, Universal, Sony, and Warner sued AI music companies including Suno, arguing their models were trained on copyrighted recordings without permission. The suit lands on both sides of the border. It is still unsettled whether AI-generated songs carry copyright at all, and who owns it when they do. Chinese platforms have responded with guardrails — flagging likely impersonation of real artists — but the underlying question of training-data rights remains open everywhere. ## Honest limitations The user and song counts above come from company statements and Chinese state-media reporting, not audited figures, so they describe momentum rather than verified scale. The technical specs — 80 seconds at 44.1 kHz stereo, the 400-billion-parameter base — are drawn from Kunlun's own announcements and should be read as vendor claims. AI singing voices, while improved, still sit a noticeable step away from a real performer, especially on sustained emotion. And the copyright landscape is moving; what is permissible today may not be next year. ## What readers can do now - If you make Chinese content, try 海绵音乐 and 天工SkyMusic with the same prompt and compare Mandarin clarity directly. - If you score video, test 天谱乐 by feeding it a clip and judging whether the auto-matched music saves you a licensing search. - If you publish generated music, read the platform's terms on voice impersonation and copyright before distributing anything commercially. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-aigc-ai-music-china Free to quote with attribution and a link to the original.