[November 2025] Dango AI Development Log
Dango AI · 2025-12-04 Coming soon
[November 2025] Dango AI Development Log

Long time no see, everyone🧐

In November, we focused mainly on developing Lossy-Audio Fixer. As our October development report explained, after restarting development of this feature and achieving substantial results this time, we decided to shift the team’s development focus to this algorithm.

Although we have reached a new level, two key problems still stand in the way:

Energy Matching

As mentioned before, if we treat a spectrogram as a kind of image and ignore its specific meaning, its high frequencies represent texture, such as instrumental harmonics, while its low frequencies represent the envelope, such as an instrument’s noise floor, shape, and approximate energy.

Last month, we reported that our new algorithm could already generate convincing high-frequency content. But that is not the end: how loud the generated content should be—in other words, its energy—is still a challenge.

We view lossy audio restoration as a standard inverse problem. An inverse problem has an oracle solution, but most inverse problems destroy the means or route to finding it. For example, suppose EQ distorts an audio file, making high-frequency energy very strong in one region. Lossy encoding discards high frequencies, removing that region. If the EQ also had a particularly narrow Q value there (Q describes how wide a frequency range around an EQ point is affected), the network cannot find clues in nearby frequency bins and can only output a statistical average. That clearly does not match the oracle solution.

Although our network can now output meaningful content, it lacks energy-matching capability, so the output energy may not match the lossless original. We are experimenting with solutions, which may take time. If results remain unsatisfactory for too long, Dango also has a fallback: give users editable energy matching. By default, Dango would output an approximate result for beginners to repair in one click. Advanced users who find that insufficient could freely edit the output energy, repair strength, and even the repair “height,” meaning frequency. These are the main issues we will address this month, and we will report experimental progress next month.

Phase Matching

Compared with energy matching, phase matching is a relatively small issue🧐. In regions where lossy and lossless audio overlap, most lossy encoders preserve fairly accurate phase information that we can use. We only need to focus on phase in the regions requiring restoration.

Our first-generation algorithm will not focus too heavily on phase. We apply light constraints and simply aim to do no harm, as long as phase does not affect amplitude or listening quality. This may mean Dango’s output scores relatively poorly on metrics, such as everyone’s favorite SDR🤓☝️. That is fine. Dango has always approached things from an industrial and practical perspective. We give greater weight to real human listening preferences than to metrics.

What Comes Next?

The release of Lossy-Audio Fixer is now a certainty, but some problems still need solving before it reaches everyone. We will continue developing and refining it in December. If progress is smooth, it will launch in January–February; otherwise, further refinement may be needed depending on circumstances. Some users may consider this feature niche compared with Dango’s others, or a distraction from our main work. Its development will also feed into our future separation features, providing mature solutions for next-generation separation algorithms. Stay tuned🧐