How is a “good” or “bad” singer defined in signal processing terms?

I was watching the American and Indian idol auditions on youtube earlier in the day. I watched both successful and failed contestants and I was able to distinguish the really bad from the good singers even if I did not recognize the song or had ever heard the melody. Among the good singers, however, my limited knowledge of music prevented me from further discerning how good or bad they were. That set me thinking and I have the following questions to ask:

1. When we say that a singer is "good", what do we mean by it in signal processing terms?
2. What factors/characteristics should we consider when we say (with some authority) that the person is a "good" singer? Is it pitch, the "tune", "melody"? How much weight should these (and other parameters) have when taking such a decision?
3. How important a role would pitch play in this? I mean, if a man was to sing My heart will go on in perfect tune (with low pitch), would he get a score higher than a woman who sang the same song (but much less perfectly)?
4. When the contestants are shown singing, there is no background score playing. Yet, we are able to figure out who is singing well and who is not. So does that mean that the background score is pretty much irrelevant when it comes to deciding good/bad singing ability?
5. Can the above points be extended to all types of songs and genres? For instance, in most classical music (vocal), the singer is pretty much primary. The instruments etc. are present simply to aid the singer. These songs are tougher to sing (any Hindustani classical or Largo al Factotum). Songs from rock/heavy metal (We will rock you/Suicide blonde) are a little easier to sing (IMO), because the instruments are primary and vocals are secondary.
6. How do we (if at all) compare two songs?

I suppose that the above problem would eventually lead me to look at karaoke scoring algorithms, of which there are good open-source options (Performous, UltraStar Deluxe etc.). But before I look at these in detail, I thought it would be great to put this out and see what ideas others in the community (with a far greater knowledge of music and ability) have on this.

P.S: Reference to some papers/material on this would be great.

P.S: Some of the above examples/points are my opinions. Given what I know about music, they may be way off the mark. Please feel free to add/suggest more questions or correct opinions I may have expressed. I will be only too happy to make the necessary edits.

• The question is good, but I don't think it's valid in the context of this web site. It seems very broad and there isn't a single short answer that will do here. This is more of a discussion thread, not a Q&A question. I'm not downvoting, because the question is well-posed, but voting to close, because it's not valid for the scope of this site. – Phonon Sep 20 '11 at 18:11
• @Phonon: I am willing to restrict the scope of the question. Any suggestions on what will fit the scope? – Sriram Sep 20 '11 at 19:21
• @Sriram You've got maybe 50 good questions in there, each one of which could be a research topic. They're great questions, but don't necessarily have easy answers (other than maybe a link to some books). To start, I'd suggest asking about defining/calculating individual metrics on a vocal in new questions, and sticking to one metric per question. For example, "How can the start time of notes in a vocal performance be calculated?" is probably specific enough. If you think about how little of your current question that would answer, it'll give you a sense for how open-ended it is. – datageist Sep 20 '11 at 21:03
• @datageist: Ok. Let me think about the question some more and come up with some more specific questions. The entire objective of this question was to bounce different ideas off of the community (from music and other DSP domains) and not so much to look for specific answers. I know that there are no specific answers to be had when we are talking of measuring subjective experiences quantitatively. – Sriram Sep 21 '11 at 6:00
• Note to moderators: Please do not lock/close this question. I will try and find more on the topic and reduce the scope. Thanks! – Sriram Sep 21 '11 at 6:02

First order, some DSP analysis might be able to detect central pitch, pitch variation (vibrato), timbre (overtone richness), and note onset, and say whether these are grossly out-of-tune, thin, or off-time, by how much and how often.

However, second-order, good musicians play with subtile variations in pitch (equal vs. just intonation, barbershop quartet harmonization as the key changes, etc.) and in timing (jazz, blues, "swing" rhythms, etc.). How to "score" the latter is probably more a topic for computer AI than DSP.

• Not sure if this mention is appropriate for this site, but I have an iPad app in the App store that roughly scores singing pitch by displaying different colors for in-tune and way-out-of-tune notes. – hotpaw2 Sep 21 '11 at 19:46
• You can see the section on self-promotion in the faq. I think as long as it is very much on-topic to the question, blends in organically with the answer (i.e., your answer must not exist solely to plug the link), you disclose your affiliation and you don't do it every answer, it is fine to promote your app. – Lorem Ipsum Sep 21 '11 at 21:03
• @hotpaw2 - Based on your website, I'm guessing that you're referring to Sing-in tuna. However, following its link leads me to the inTune Strobe Guitar tuner. Where is this iPad app? – Kevin Vermeer Sep 22 '11 at 12:43
• As an author of a commercial app which approximately accomplishes the goal stated in the question, you're an expert on the topic. I appreciate your caution, but please don't hold back when you have something valuable to contribute! – Kevin Vermeer Sep 22 '11 at 12:46
• @Kevin - Thanks! HotPaw website link for Sing-inTuna fixed. Feedback for this app is mostly from singers using it to help them practice not singing too off-key. Which would be "bad" in terms of the OPs original question. "Goodness" is much harder to measure. – hotpaw2 Sep 22 '11 at 15:10

This is a difficult question to answer in a scientifically rigorous fashion given that music (an art form) is generally a subjective issue and people will differ greatly on opinion. The problem is that, quantitatively, a singer could be "perfect" in pitch but have an objectionable tonal quality (timbre in musical terms). This can be due to accent, punctuation, emphasis etc... Most great singers the world has never heard of fall into this category. Then there is the opposite problem, someone who has poor pitch control but a pleasing voice. Frank Sinatra, is a fantastic example of this...it widely accepted that he sang slightly flat a great deal of the time. Tom Waits and Bob Dylan are yet more examples of less than perfect singers but there seems to be a general consensus that they have pleasing voices but again this all purely subjective. In DSP terms you would be forced to stick with quantitative features of the voice signal such as pitch accuracy and note timing (rhythm). Although not a "serious" voice analysis application, the technology behind Sony's Playstation game Singstar, uses a short-time frequency analysis technique to derive pitch from the players singing voice and then compares it against a time aligned annotation of the vocal track from the song. Other technologies use frequency analysis followed by pitch shifting to actually correct pitch in the singing voice in realtime > Eg> http://www.celemony.com/cms/ and http://www.antarestech.com/.

• Thanks for the reply! I know it does not have clear answers, but I wanted to see what thoughts the community (especially those more familiar with music than me) had on this.. and also if I could be pointed to some more reference material. Your answer is helpful. thanks! – Sriram Sep 21 '11 at 17:18
• Yes, the Music Technology Group in UPF Barcelona do lots of research in the area, here is one specific publication >mtg.upf.edu/files/publications/… .They have many more related also. Go here > mtg.upf.edu/research/publications , and type singing voice into the keyword search on that page for a list of publications in the area. – Dan Barry Sep 22 '11 at 8:24

I think it is possible to make a DSP application (Android, iPhone, or whatever) that records a singing voice and rate the performance of the singer (maybe even provide feedback to the singer). One of the things that the DSP program could do is to examine if the notes in the recording are at the correct frequencies or if they are off. If the program knows the song, it might also be able to examine if the singer maintains a correct tempo.

• so for "grading" a song, tempo, notes and duration for which those notes are held are important? What about the "alignment" of spoken word with the "tune"? Would that be important too? Are there any papers that you can refer me to on this topic? – Sriram Sep 20 '11 at 19:21
• It seems pretty clear to me the the tuning of the notes and their duration must be important because if they are off, the singer must be "bad". I'm not sure we're on the same page. What I proposed is really simple. I'm not a singer. My understanding is that you want to analyze the voice of a singer and rate "good" or "bad". – niaren Sep 20 '11 at 19:37
• Don't get me wrong. I am no musician myself. It is just that when I was thinking of this problem, I came up with so many terms that I could not define in clear terms - I had only vague terms like "good" etc. to go with. I am just trying to understand this better. – Sriram Sep 20 '11 at 19:41
• While you may be able to assert that someone who is off-pitch is "bad," I would caution against doing the opposite. There is a human, qualitative part of measuring singing quality that I think is difficult to measure quantitatively. One illustration: the "Autotune" technique that is used often in modern American pop music. It uses signal processing to force the notes to be on-pitch, but many people would tell you that it is not terribly pleasing to the ear. – Jason R Sep 20 '11 at 19:51
• @niaren - No disrespect intended, but this answer seems very weak. The possibility of creating a "DSP application" on a mobile device is completely irrelevant to the question How is a “good” or “bad” singer defined in signal processing terms? Removing this sentence reduces your answer to the criteria of pitch and timing, postulating that a program could "examine" them without defining how that examination would work. It has some potential, but I don't think it deserves the upvotes it's earned. – Kevin Vermeer Sep 22 '11 at 12:37