![[February 2025] Dango AI Development Log](https://tuanziai.com/content/images/2025/03/--rnd.jpg)
Our development team focused throughout February on restoring lossy audio quality. We first planned this feature around two years ago but had never implemented it. Our entire lab is now working on it at full strength—we counted roughly 700 experiments over February.
Unfortunately, although we split the work into several directions and ran different experiments in parallel on different machines, the results remain unsatisfactory 😭
Lossy and lossless songs differ little in the time domain (waveform); the main differences are in the frequency domain. High frequencies usually look cut off entirely, with hole-like defects in the mid and low frequencies.
There is currently no similar product on the market, though a few related works exist in academia. The closest to our task is Apollo, a paper by prominent researchers in China on restoring lossy audio. Although it is impressive research, our evaluation suggests Apollo is unsuitable for this task for these reasons:
- Horizontal-line artifacts abound, turning one kind of lossy audio into another. We believe embedded energy information during spectrogram-to-feature conversion causes this.
- Mid and low frequencies become blurred and averaged out, which we believe results from the transformer architecture.
- High frequencies cannot be restored. Apart from easy-to-learn full-height patterns such as drums and vocal sibilance, no meaningful energy is generated.
We therefore decided to build our own architecture for this task. It already solves issues 1 and 2 perfectly, producing clear output without artifacts.

In this image, Dango’s output spectrogram is clearer and artifact-free. Even some high frequencies are restored, apparently recovering part of the breathiness in vocals.
This is still far short of our expectations. We want meaningful high-frequency restoration. Vocals appear to have some high-frequency content, but instruments reveal the weakness:

This is a simple string instrument sound with a simple shape: harmonics occur at regular frequency intervals above the fundamental. Dango seems to predict some blurred harmonics, while Apollo makes do with averaged, blurry artifacts. But even Dango’s result is unusable. Near the high frequencies, disconnected harmonics look like grains; our algorithm cannot join these grains to restore the harmonics. Its high-frequency predictions are also uncertain: faint shapes appear, but no meaningful energy lines are generated.
The first two problems were relatively easy for us to solve. We remain stuck on the third: restoring meaningful high frequencies. We have tried many approaches without clear success.
After discussion, we decided to add another month of experiments. We have some new directions and will keep looking for improvements throughout March. If there is still no significant progress, Dango may abandon development (leaving it to the academic experts 🧐), as the current results are nowhere near practical use. We would then focus fully on the 11.0 separation algorithms. Thank you for your patience 😥
Also, if you are an expert researching a similar area, you can contact us. We welcome your solutions and are willing to pay.















