Generating Guqin: Four AI Music Tools, One Honest Verdicteric

When we built the Chinese poetry feature in Tyokiie, the ink-wash art came from a small GPU in the next room. The music was the other half of the problem — and it sent us on a small bake-off worth writing down.

We needed a very specific thing: solo guqin — the seven-string zither — sparse, slow, meditative. No drums, no string section, no cinematic swell. Just the wood and the silk, the sound of an empty mountain studio.

That turns out to be a hard ask for AI music. Guqin is a niche timbre with a particular articulation — the slides, the harmonics, the long decays — and most generators have heard far more lo-fi and EDM than Tang-dynasty court music. So we ran the four best options we could get our hands on through one identical brief. This is the field report.

The brief

Every tool got the same intent: solo guqin, sparse and slow, meditative, no percussion, generated per mood — still (winter/snow), flowing (water), autumn (leaves), and mountain (general). Roughly 45 seconds a track, to be looped under the reader. Three things mattered, in this order:

1. Authenticity — does it sound like guqin, or like a synth pad wearing a hat?
2. Commercial rights — can we publish it on a monetized channel without a licensing landmine?
3. API access — can we automate it inside our render pipeline, or is it click-and-download?

MusicGen — great sound, wrong licence

Meta's MusicGen (part of AudioCraft), run locally on the same RTX 3060 that paints our landscapes, surprised us. The medium model produced genuinely pleasant guqin — believable plucks, reasonable phrasing, a real sense of space. On pure sound quality it was the best of the open options by a clear margin.

The catch is the licence. The MusicGen model weights are released under CC-BY-NC 4.0 — non-commercial only. The code is permissive; the thing that makes the sound is not. For a hobby project that is fine. For a monetized channel it is a non-starter — beautiful results we simply cannot ship.

Stable Audio — not there yet for guqin

Stable Audio Open runs locally too and is a capable general music model. But on the guqin brief specifically it could not hold the timbre — the plucked-string character drifted toward generic ambient pads, and the articulation that makes guqin recognizable just was not there. It may get there; today it does not.

ElevenLabs Music — the one we shipped

ElevenLabs Music landed in the sweet spot. Quality is on par with MusicGen — arguably a touch better on our brief — and crucially it clears both practical hurdles at once. Commercial rights come with a paid plan, so the generated audio is ours to publish; and there is a real API (POST /v1/music), so track generation drops straight into the render pipeline, mood-keyed, with no manual export step.

That combination — good-enough sound, commercially clear, and scriptable — is why every backing track in our poetry videos is now ElevenLabs Music.

Suno — the best, and the one we can't use

And then there is Suno. On raw authenticity it is the closest to the real thing we have heard from any generator — the kind of output where you stop A/B-testing and just listen. If the only question were "does this sound like a master playing guqin," Suno wins.

But there is no public API. Everything is through the app, by hand. For a pipeline rendering dozens of videos, hand-generating and downloading each track breaks the whole model. It is a genuine shame: the best voice on the instrument is the one we cannot wire in.

The scoreboard

What we learned

The best-sounding model and the right model are rarely the same one. MusicGen and Suno both beat our production choice on some axis — sound, authenticity — but a soundtrack pipeline lives or dies on the boring stuff: can you licence it, and can you automate it. ElevenLabs won by being good enough on every axis instead of best on one.

If Suno ever opens an API, we will re-run this bake-off the same afternoon.

You can hear the result in the ambient poetry videos inside Tyokiie — every backing track is original AI-composed guqin, generated with ElevenLabs Music under a paid plan (commercial-use rights), not sampled from any recording. The poems are public-domain classics; the ink-wash art is our own local FLUX pipeline.

Next in this series: stitching the painted landscape, the brushed-on verse, and the guqin track into a single looping video — and what Google Veo can and can't do for a still, meditative scene.