3
$\begingroup$

I have a fairly large collection of audio/video bootlegs recorded at live performances by various artists. The recordings all come from very different sources and people through decades, and all of the recordings in the collection are just audio/video files. Since the collection is large, the total running length of the collection is probably hundreds hours.

The quality of each recording varies: some of the recordings sound excellent (let's call it "10"), but some of them sound more like noise ("1") remaining a point of interest for a particular artists collector. Let me say this way, by "quality" I mean something the way "how good" it looks and sounds to me or anyone else (and not the audio/video codecs settings the recordings are encoded or compressed with; and not how a particular artist performs -- it's a matter of personal taste, not audio/video quality). When I started collecting bootlegs more than a decade ago, I remember, I could find characteristics for many of them like "Quality: A+" (really nice), or "Quality: B-" (not that bad, but makes interest to a collector.)

For example:

  1. if a live recording sounds as if it is recorded in a studio, then it might be evaluated to "9" or even "10" (regardless if either a lossy or a loseless audio-codec used);

  2. if a recording shows up noticeable visual VHS artifacts like blue/red/green stripes (just because the recording was recorded on a tape), but the overall picture is pretty good, it might be evaluated say to "5+" up to "7";

  3. if the recording sounds very "bassy" and low frequencies heavily prevail the high ones, it might be evaluated to "3-" since the audio might be considered very low quality, etc. If such a thing exists, I guess it might also to be applicable for audio, video and images;

  4. and more...

Is it possible to analyze a recording in a software way, not listening/viewing to it, to "determine"/"evaluate" its subjective quality?

This question looks pretty much similar to Analyzing the quality of a music track (and probably audio quality evaluation ) , but can't really tell how close it is.

(Please note I have zero knowledge in this area, may use wrong terms and may ask something unreal. The only reason I'm trying to find it out is editing the files metadata by putting the "quality" tags into it, thus evaluating the average/overall quality of the entire collection not spending weeks of listening or watching to all of the recordings regardless hardware I might use. Also not sure if the question is better to ask at Software Recommendations or Sound Design though.)

$\endgroup$

3 Answers 3

2
$\begingroup$

Is it possible to analyze a recording in a software way ... to "determine"/"evaluate" its subjective quality

Yes, very possible; all one needs is to define, mathematically, what "good quality" is - and enough data. The full pipeline may involve:

  1. Understanding basic signal processing decomposition of audio (i.e. building blocks) - see DSP Guide
  2. Extract relevant features with a transform, e.g. wavelet scattering, time-frequency scattering, synchrosqueezing, MFCC, etc.
  3. Apply learning algorithm with suitable objective function:
    • The function can be a measure of "distance" in feature space. This can be done by having a "template" for what's considered "good music" based on extracted features.
    • Test phase can involve direct subjective assessment of test subjects, as in Natural Language Processing - but the optimization function must be entirely mathematical.
    • Success requires sufficient data. The requirement is lowered with better feature engineering and transfer learning (e.g. NLP transformers).
$\endgroup$
2
  • $\begingroup$ "all one needs is to define, mathematically, what 'good quality' is" - To me this sounds as it's underestimating the fundamental challenge of the task drastically. There have been numerous attempts to find a mathematical description of perceived quality - with good results, that, however, in my opinion still are far from optimal. This is taking into account that most of the developed methods are even based on knowledge of a reference signal, to which the perceived quality of a test signal is compared. $\endgroup$
    – applesoup
    Commented Aug 26, 2021 at 18:11
  • $\begingroup$ @applesoup Agreed, it's much easier said than done. But the challenge is similar to NLP; perhaps with equal research effort we'd get same or better results. We already have AI-composed music, for one - presumably there's far less interest in music evaluation systems. $\endgroup$ Commented Aug 27, 2021 at 6:32
0
$\begingroup$

I think if there was really very much to the latest "AI" craze, doing what you want here would be a piece of cake. I don't think it's doable at all. Good luck

$\endgroup$
0
$\begingroup$

What you are looking for, is a perceptual metric targeting purely appeal. No such thing exists yet, and for good reason. The problem with creating such a general metric for appeal is that appeal is always also influenced by assumed fidelity, the difference to the assumed state of the original, and also highly subjective. For such a metric to be somewhat robust, we'd need an extremely huge dataset of very diverse samples, a wide variety of distortion applied and multiple people scoring the quality for each combination, and then hope that you get a useful metric with machine learning. This metric in the end would not even be useful for next generation codec development, too easily exploited by new compression methods. And the main focus of any encoder should be fidelity anyway, not appeal, although appeal should generally play more of a role the further you move away from transparency.

However, when only targeting the quality loss that happened during encoding in a specific old simple format, preferably widely used, like JPEG, H.264 or MP3, I can see creating an appeal metric as viable. Especially in the case of JPEG, where the coding tools available are so sparse that it should be much simpler to create a somewhat robust appeal metric, which in this case would target JPEG compression artifacts.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.