- Joined
- Mar 6, 2021
- Messages
- 214
- Likes
- 289
The only thing that machine-learning needs is training data. Additional constraints can also be added, if desired. Since the resultant algorithm runs on GPUs it is very fast and efficient.
-One way to begin would be to pick up something like a M-scaler and use its input/output as training data. That's assuming that the M-scaler actually does a good job of it.
-Another approach would be to begin with the best quality audio recordings one could find, then downgrade them (in a variety of ways, downsampling, bit compression, etc.). Then the downgraded files would serve as input and the original files would be the output.
What kinds of constraints could one add? It is important to understand the entire audio chain, so calibrating sound output relative to file input to achieve the ultimate result at the listening position should be kept as a first priority. In fact, a GPU-based DSP could be designed that takes care of room corrections AND upscaling. Running the calibration would be more time intensive, and could take a while to compile the resultant ML algorithm. Then there is the biophysics of the audio reception, from ear dynamics to neurology.
Anyways, all of this is already possible to do...surprised nobody has done it. Than again, I hear that big outfits like MS and Sony have been working on this...so it is coming around, sooner or later. If somebody were smart enough to figure out a good process and patent it, those giant corps would be eager to buy it from them.
-One way to begin would be to pick up something like a M-scaler and use its input/output as training data. That's assuming that the M-scaler actually does a good job of it.
-Another approach would be to begin with the best quality audio recordings one could find, then downgrade them (in a variety of ways, downsampling, bit compression, etc.). Then the downgraded files would serve as input and the original files would be the output.
What kinds of constraints could one add? It is important to understand the entire audio chain, so calibrating sound output relative to file input to achieve the ultimate result at the listening position should be kept as a first priority. In fact, a GPU-based DSP could be designed that takes care of room corrections AND upscaling. Running the calibration would be more time intensive, and could take a while to compile the resultant ML algorithm. Then there is the biophysics of the audio reception, from ear dynamics to neurology.
Anyways, all of this is already possible to do...surprised nobody has done it. Than again, I hear that big outfits like MS and Sony have been working on this...so it is coming around, sooner or later. If somebody were smart enough to figure out a good process and patent it, those giant corps would be eager to buy it from them.
