![[September 2026] Dango AI Development Log](https://tuanziai.com/content/images/2026/10/------rnd--2-.jpg)
( ̄ω ̄( ̄ω ̄〃 ( ̄ω ̄〃)ゝFirst, happy National Day to everyone~
This month—and the coming months—remains a time for development🧐. Let’s look at September’s progress.
Vocal Remover 12.0 in Development
This month, our main development work remains focused on architectural design for Vocal Remover.
Unlike before, the new architecture is almost 3–4 times larger than the previous generation, meaning close to 1 billion parameters. This is a major engineering and experimental challenge🤒. Each failed experiment costs more, and experiments take longer.
We divided the algorithm into three abstract goals. All three must be achieved to complete this development cycle.
We are still in the mid-to-late stage of goal 1. This month’s progress is a little embarrassing🤒. We spent nearly 15–20 days experimenting in a completely wrong direction, losing a great deal of time. We have now changed direction.
(Short version: all that buildup, and nothing to show for it.jpg)
Compared with last month, however, we largely settled the training objective in the latter part of this month. The current model is better than the previous generation on most songs, but some issues remain unimproved. We are fine-tuning specifically for those issues to obtain more consistent, accurate output.
On metrics alone, we reached an SDR score of 18.83 dB last month and 18.89 dB this month. In addition to automatically calculated metrics, we have an internal human listening-test platform. Our output quality this month is much higher than last month.
New Minor Noise Removal Model (Working Name) Confirmed for Release
While developing 12.0, we also allocated some computing power to experimenting with this new model.
As introduced in last month’s log, this model further and thoroughly removes tiny noises, background hiss, headphone-monitor bleed, and other very quiet but still audible, irritating sounds from recorded vocals.
We are very satisfied with the final result🧐. We compared open-source implementations and those in some software and websites. Most solutions only extract vocals from songs rather than specifically removing these noises, so their output may not change at all—the quiet sounds remain. A few solutions can partially, but not completely, remove some sounds.
The reason is that these models’ architectures are not sensitive enough. On the engineering side, completely noise-free vocal training material is also very difficult to obtain. Training on ordinary vocals limits the network’s ceiling. However much training it receives, it will always ignore certain sounds.
Dango tackled the problem from both model and engineering perspectives and solved it thoroughly. The final output meets our release standards, so this model will launch as one of the Vocal Remover series algorithms alongside 12.0. This provides two solutions for vocal extraction: extracting vocals from noisy music or sound environments, or further post-processing master vocal recordings to remove unwanted content.
(From top to bottom: original audio, Dango-processed audio, and the noise extracted by Dango.) Since this model removes extremely quiet sounds, the difference may be imperceptible through speakers. For listening tests, we recommend headphones and paying attention to the volume. The audio comes from an intermediate model and does not represent the final result.
What Comes Next?
The Minor Noise Removal model is close to satisfactory. We have paused its final fine-tuning and allocated all training computing power to developing the 12.0 Vocal Remover series, trying to accelerate our experimentation and development.
We estimate the 12.0 series is currently only about 25% complete. There is no need to extrapolate its eventual release date linearly from the past few months. Like in a game, a favorable event one day—“a villager shared nuclear fusion technology with you.jpg”—could cause development progress to jump significantly👻
Next month, in theory, we will nearly finish the work on goal 1 of our three abstract goals, hopefully reaching a qualitative conclusion about the 12.0 series DangoNet6 architecture.















