Play a generated Mandarin ballad to a native speaker and watch where they flinch. It is rarely the chord change. It is the moment the singer swallows a syllable, hits the wrong tone on a character that has two readings, or stretches a word until the line stops meaning anything.
That flinch is measurable. In music AI it is called phoneme error rate (PER) — roughly, how often the sung audio diverges from the intended sounds of the lyric. It is the least glamorous number in the field and the one that decides whether a generated track can be used in a real release rather than a demo reel.
Tencent AI Lab's SongGeneration 2 reports a PER of 8.55%. For comparison, the same reporting puts Suno V5 at 12.4% and Mureka v8 at 9.96%. Those are vendor-reported figures from listening and alignment tests, not a third-party leaderboard, and I treat them accordingly. But the direction is worth taking seriously: a Chinese open-weight music model is now claiming lyric accuracy in the same band as the leading commercial systems, and slightly ahead.
What the model actually is
SongGeneration 2, also published as LeVo 2, takes tagged lyrics — [verse], [chorus], [bridge] — and produces a full song with vocals and accompaniment, up to four and a half minutes. It handles Chinese, English, Spanish and Japanese lyrics. The architecture pairs a language model that plans the song's structure with a second stage that predicts vocal and instrumental tracks in parallel, plus a diffusion-based audio codec.
Two details matter more than the parameter count. First, the weights are open. The large checkpoint runs on roughly 22GB of GPU memory, which puts it on a single high-end consumer card or a rented hour of cloud compute — not a datacentre. There is also a Fast variant on Hugging Face that trades a little fidelity for speed. Second, it accepts a reference clip as a style prompt, which is how most working producers would actually use it: not "write me a song", but "sound like this, with these words".
Hosted inference is cheap enough to make the economics concrete. One third-party playground charges about US$0.0006 per second of generated audio, which works out to roughly US$0.16 for a four-and-a-half-minute song (≈HK$1.25). At that price, generating twenty takes to get one good chorus is a normal workflow rather than a budget decision.
Why Chinese is the harder test
Western music models were trained on a language where stress carries a lot of the melody. Mandarin carries meaning in tone: the same syllable at a different pitch is a different word. A model that treats the vocal line as a melodic contour and the lyric as a sequence of syllables will produce something that sounds like singing and parses as nonsense.
This is why the phoneme number is a bigger deal for Chinese users than for English ones, and why the domestic competition has organised around it. Song generation (歌曲生成) is the product; singing voice synthesis (歌声合成) is the craft underneath it, and the craft is where tones live. The claim is not that Chinese models write better songs. It is that they mispronounce fewer words, which is the difference between a track you can post and a track you have to re-record.
There is a second, less discussed layer: polyphonic characters (多音字). A character with two readings is a coin flip that the model has to resolve from context — and song lyrics deliberately break context for rhyme and metre. English models handle this badly because the training signal is thin. It is the kind of problem that only gets solved by people who care about the language, which is an argument for local models that has nothing to do with nationalism and everything to do with data.
What open weights change, and what they do not
The practical consequence of an open checkpoint is not free music. It is that a small studio can fine-tune on its own catalogue, keep the outputs on its own machines, and build a product without paying per-second fees to a foreign API. For a Chinese games studio that needs forty minutes of background score a month, or an audiobook publisher that needs a theme song per title, that is a real change in the cost structure.
It also changes the licensing conversation. A generated song built on a model you host is a different contractual object from one generated on someone else's server under terms that can be revised. That distinction is doing more work in procurement departments right now than any quality benchmark.
What open weights do not fix is the part that still fails: the hook. AI can produce an arrangement, a competent melody and a clean vocal. What it tends to produce is a chorus that is pleasant and forgettable — the melodic equivalent of a stock photo. Publishers who have listened to a lot of this output describe the same thing: the verse is fine, the bridge is fine, and nothing makes you want to hear it again.
Where the Chinese market sits
The competitive logic in China has shifted from quality to access. NetEase's 天音 (Tianyin) studio has been pushing a free, one-tap songwriting mode that keeps copyright with the user on condition of an AI label. Tencent went the other way and gave away the weights. Both moves assume the same thing: the model is not the moat, the workflow and the rights position are.
For a global reader, the useful takeaway is not that Chinese music AI has caught up. It is that the measurement that decides commercial usability — lyric accuracy — is now public, comparable, and being competed on. That is a healthier situation than a field where everyone argues about vibes.
Honest limitations
The 8.55%, 12.4% and 9.96% figures come from Tencent's own reporting and a listening study it commissioned; I could not independently reproduce them, and PER is sensitive to how you segment and align audio, so cross-vendor comparison is approximate. I have not heard a controlled blind test of SongGeneration 2 against Suno V5 on the same lyrics. Pricing cited is one third-party host and will vary. I have not addressed training-data provenance, which is the unresolved legal question across the entire category.
What readers can do now
- If you are choosing a music model for Chinese-language work, run your own ten-line lyric test and count the mispronunciations yourself. It takes twenty minutes and beats any benchmark table.
- If you are buying generated music for commercial use, confirm in writing whether the licence travels with the weights you host or only with the service you rent.
- If you are building on an open checkpoint, budget for a fine-tune. The base model writes a competent song; your catalogue is what makes it sound like yours.
