This thread made me think, and sent me off on a tangent…
I have written a crude prototype program that plays a series of notes (through headphones preferably), and detects the input from the microphone. It analyses the input to tell you when you are playing the correct note on your whistle.
I thought maybe that this could develop into a learn-by-ear tutor. For example it could play a three-note sequence over and over until the user got it right three times in a row, then move on to a four note sequence.
The basic technology is there - it needs some direction and window dressing. What think you?
It has been done, not too well, several times in software intended to teach exactly what you describe. I think a new attempt would be a good thing. Go for it!
In the best of all possible worlds it would be good to make the program independent of the sequences and let the sequences be developed either by a set series of patterns created by the user or be a set of rules to decide about which note patterns are acceptable. The pitch detection and accuracy part of the program are the easy part (though of course you have to decide how accurate the singer of player needs to be). The choice of sequences or rules for creating sequences are the part that makes such a program useful. And, they are truly difficult to decide upon. In my years as an active member of the Association for Technology in Music Instruction (ATMI) that was always what members fastened upon and disagreed about.
I think it would be nice to develop a series of sequences based on common patterns in some tradition or another. Irish tunes certainly suggest that sort of thing. So do bluegrass and old timey traditions. Blues too… Well you get the idea.
The fact that it has been done before is good (I am not mad) and bad (I have competition ).
As far as the note sequences that the user had to emulate, I was going to allow run-time specification in pseudo-abc code. I see that part of the program as a simple (but potentially tedious) implementation of a partial ABC parser.
What happens with the fragments next is more complicated. For example, should the program slowly build up an entire tune starting with a few notes and increasing fragment size as the whistler gets more competent? I think not. This should not endeavour to be a tune-teacher. Tunes should not be learnt from MIDI output.
My aim is to create software that will let me practice “hear a pitch - play that pitch”, i the hope that I will fill i on a skill that I am sadly lacking. I would therefore think that short(ish?) sequences of notes would be better than a whole tune.
In some ways easy - though there are constraints. I cannot get response below 100ms for the bottom end of low D (D4) or 50ms for D5. I think this may be an absolute limit, but my knowledge in this are is thin. (For those wanting to help, I am performing a Discrete Fourier Transform, and then pulling out the first frequency peak exceeding 50% of the amplitude range).
I then find the best match for the frequency in a 12-tone equal tempered scale. I could constrain the search to the subset of semitones available to the whistle, making the match more forgiving, but this did not seem to improve accuracy much. It might help if I want to decrease the response times.
I could display the degree of divergence from the ideal frequency, but I am not sure how much use that would be. For me personally that would be too much information.
What constitutes a ‘correct’ note? Hugely difficult question. I was thinking of calculating the amount of time that the correct pitch was played (by the user) , and expressing it as a fraction of the time the whistle was played by the program. If this fraction is set too high, we are back to teaching ‘tunes’, as the fraction will be improved by repetition. If we make the fraction to low, we might reward random success too highly.
You are a very useful commentator on these matters then. My attitude to these differences is to ask ‘what are the meta rules?’. All the different rules that people argue over probably have some common themes and patterns to them. I then implement the common themes, and allow individual users to select the specific implementations of the rules. Unless of course it looks as if I am implementing ‘yet another language’ whereupon I pass the problem off to live javaScript entry (See TWJCalc for an example).
If anyone wants to play with a (very) primitive prototype let me know. It currently only teaches a scale from D5 to D6, to prove the generate-hear-play-detect-assess loop.
I didn’t look in detail at “World’s Best Ear-Trainer” but it seems to illustrate some of the issues that show up in this sort of program. Please be aware that I am not being critical of that program, but rather using what, in my rather cursory examination, it appears to do as examples of some of the issues. I suspect this discussion is getting rather off track for this forum, and invite Phil to contact me off list for further thought I might have.
!) It appears to use randomly generated “melodies.” If they are truly random they may actually not be particularly well centered in a tonality and so remove from the user the possibility of hearing the melody by relating it to a tonal center. In most traditional music there is some sort of (usually strong) tonal center which is very helpful in learning to hear melodies. Remove that and you make things more difficult and perhaps less relevant to the learner. And, because the sequence of notes is random you may have no sense of tonal center at all even when choosing to use pitch sets that imply one. Using only a few pitches this is not an issue. It becomes an issue when you have an entire scale of notes.
It seems to generate melodies of some set length rather than melodies that “grow” as the user gets them correct. That is X correct responses to 3 note melodies moves you to four note melodies. Of course the user can manually make that change at any time, but some folks will argue that the program should move you forward (or backward) through some sort of logical sequence.
Melodies based on a single note duration are, of course, not particularly melodic. “Real” melodies aren’t based on a single note length of course, but adding a rhythmic component to a random generator is quite difficult. It was difficult for me to see from the examples I looked at whether this particular program does a good job of that thorny problem.
Because the generated melodies are random they will not help reinforce any particular style of music.
Because the randomization appears to based on note usage rather than on interval usage many folks who teach ear training of this sort will reject it. They may prefer something that builds, as the Kodaly approach does on simple patterns: Sol-Mi (say G-E) patterns become La-Sol-Mi (A-G-E) patterns which do not allow the La-Mi (A-E) at least at the first level etc. etc. The program under discussion can do some of this, but not some aspects of it (like not allowing La-Mi but allowing La Sol Mi patterns without that jump).
The program does not appear to allow the user to hear a starting note or a tonal center note before the example is presented. It is tough to grab a pattern that could start on any of, say, 8 different notes. And conversely repeated use of a single key center can eventually lead the user to use good pitch recall from previous examples as the means of “solving” the current example. This may or may not be a good thing. Teachers will disagree. There are various options for adding this sort of capability: allow a starting pitch to be played at some point before or after the pattern is played; Allow a “tonic note” to be played before or after the pattern is played, and Allow the complete pattern of notes to be played before any patterns based on those notes are played. The visual aides of the program certainly help with this issue though.
These issues will lead some folks to reject this program completely. But, truth to be told I rather doubt there is sufficient hard data to accept or reject its approach.
I built a couple of different melodic dictation programs years ago (on the Apple II for heaven’s sake!) and used four different approaches at different times: a) A user (or teacher) prescribed group of acceptable notes and the program presented them randomly (much like the program discussed above); b) A user prescribed set of intervals constrained to the notes of one particular scale; c) Melodies were based on an algorithm presented in an article (Scientific American?) that discussed how to create “banal tunes,” and d) Instructors created “melody banks” where specific melodic patterns of various lengths were grouped together and randomly used. Learners were moved through the series of groups of patterns based on competency rules (the rules could be modified too). From my perspective c and d were the most successful and useful. But, they place the greatest demands on the program creator or the instructor too since the needed series of patterns (which must be quite large to avoid users memorizing the examples) has to be created and organized as to difficulty. I mention all of this not because I think I found the key to all of this, but rather to indicate some of the issues.
But, we’re a long way from whistle issues here. I suggest that either Phil and I talk by PM or perhaps this topic should be moved to the “technology” forum. In any case this is my last post on this topic for that reason. I’m fascinated by the issue, but I truly think this is the wrong forum to discuss it.
I think some of what the OP wants will come from just playing the whistle. I think that, over time, your brain learns to correlate what you are hearing with finger positions on the instrument and some of the skill just falls into place. Targeted ear training no doubt moves things along faster.