How can I clean and isolate audio?

[lmv]

I was about to give up on converting audio to MIDI when I had this idea: what if I clean the audio file to remove all noise and leave just the acoustic guitar sound? That should be possible. It doesn't have to sound "nice." The existing audio track doesn't sound very "nice" anyway. I just need to make sure that nothing is left but some part of each note sound bite that some software can use to identify the note. But I'm not really sure where to begin. Should I use an EQ? Would IQ4gui help me with that? An EQ works on frequencies. I think that is not what I need. Or is it?

Assume this is a note:

---x-x---o-oxo-OOOO-o-xo--ox----x-x-

Dashes and x'es are garbage I want to get rid of. Small o's are legitimate note sounds, but of poor quality. Capital O's are legitimate note sounds of good quality. Is there anything I can do separate the wheat from the chaff? I know quite a few tricks to sharpen/enhance images, but not sounds.

[thecontrolcentre]

I have found Supertone Clear very useful for removing unwanted noise, especially on field recordings. There's a demo you can try here: https://product.supertone.ai/clear

Another one to try is Mashtactic by QuickQuack : https://quikquak.com/Prod_MashTactic.html

Mashtactic was initially designed as a mashup tool. It can separate different parts of a full mix, manipulating different frequencies and stereo locations. Up to eight zones can be created that can then be panned, emphasised or cut out completely. It can be viewed as an eight band filter that exists in stereo ranges as well as frequency.

[mjolnir]

I forget which one it is, but there's a Cockos REAPER plugin that can be used for noise reduction too.

[BertKoor]

my 2 cents: if an Audio2Midi first requires you to clean up noise, either the recording is worse than a 78 rpm phonograph, or the detection algorithm is subpar.

Note that run-of-the-mill algorithms of this type work best on slowly played monophonic material. It has its reasons that midi guitars have a sensor mounted per string, and proper tracking is depending on playing style.

Find a volunteer to transcribe that little piece of music you have.

[donkey tugger]

With guitar you're always going to get lots of spurious detection. A lot of it can be attenuated by filtering out very short and very low velocity notes in your resulting midi using whatever tools your sequencer has (FL studio has excellent scripts for this, as does Reaper - sure others do too)

In terms of getting the optimum initial result, cut the low end from your sample (anything below about 100hz for acoustic guitar will end up confusing matters I've found) and also the top end a bit.

At the end of the day tis a very complex instrument with a lot of human interaction, so you need a bit of effort to get good results.

[lmv]

my 2 cents: if an Audio2Midi first requires you to clean up noise, either the recording is worse than a 78 rpm phonograph, or the detection algorithm is subpar.

It's somewhat better than a 78 rpm phonograph. But all converters I have tried fail badly. "Artificial intelligence" is the biggest misnomer of our time. There is nothing "intelligent" about it. Where I, human, hear 12 notes, the AI hears at least 37. It is 100% unable to filter noise out and by "noise" I mean even natural harmonies of the strings.

I suspect the problem partially stems from overshooting. It seems to me they are all obsessed with chords. Because chords are hard, and detecting chords is a great feat for a developer to be proud of, right? It does really seem they don't even conceive of single notes. It's like single notes don't even exist. So they try hard to turn all the sounds they "hear" in each amount of time into some chord. The result then is an ugly mess.

[Ah_Dziz]

Most polyphonic (and monophonic) pitch detection has settings for different instruments. So if you set it for "acoustic guitar" for instance, it should only be looking for up to six notes and all of those within the range of a guitar. There should also be threshold for noise at the analysis stage. I don't know what software you're using so maybe not.

If you're trying to get a solo acoustic or electric guitar recording into midi you could probably find a few guitarists that could figure it out for you.

[donkey tugger]

Most polyphonic (and monophonic) pitch detection has settings for different instruments. So if you set it for "acoustic guitar" for instance, it should only be looking for up to six notes and all of those within the range of a guitar. There should also be threshold for noise at the analysis stage. I don't know what software you're using so maybe not.

If you're trying to get a solo acoustic or electric guitar recording into midi you could probably find a few guitarists that could figure it out for you.

Good point. Given the harmonics and all that stuff generated then sometimes detection will throw out notes in the wrong octaves, which you can then limit to within the correct range using editing tools in your sequencer, or suchlike.

[lmv]

Most polyphonic (and monophonic) pitch detection has settings for different instruments. So if you set it for "acoustic guitar" for instance,

None of the converters I have tried has anything such as "acoustic guitar mode." What converters do you know that do that?

you could probably find a few guitarists that could figure it out for you.

Not really. Making friends and influencing people have never been my strongest suits. And I'm old. I haven't been out on a Friday/Saturday night in 10 years. I live in a van by the river.

[BertKoor]

Well, you never asked...

[lmv]

Like what? Do you want take a stab?

https://0x0.st/XlFj.mp3

Code: Select all

A.........................D#°
Bm.............. E7	 
A...............F#m
................B7.......E7 

A.....................  D#°
Bm.................... E7 
A.....................F#m
.............D7..........E7

A.................... D7°
..............Bm.........E7
F#.....................Bm
........F#°...............E7

A..................D7°
.............Bm......Dm
A..................F#m
............B7........E7

I appreciate any help. Would also be interesting if someone managed to extract a decent MIDI rendition of it and let us know what software did the miracle. I'm going for 58 BPM on that one.

[BertKoor]

Like what? Do you want take a stab?

Working on that...
The audio is somewhat messy indeed. Recorded in a pub, open stage night?
I'll respond in your other thread which is about the midi itself.

[donkey tugger]

This was the best I could get it;

MP3 using a sampler patch I made of one of my acoustics;

https://drive.google.com/file/d/1-8M0t2 ... sp=sharing

Midi file;

https://drive.google.com/file/d/1-5yoQT ... sp=sharing

Firstly eq'd the file quite drastically;

actest.jpg

Used Neural Note with these settings;

nnote.jpg

Then in FL studio used the limit function in the piano roll to ensure the notes stay within guitar range, and finally used the select notes script to delete very quiet and very short notes which would generally be spurious detections.

Not perfect by any means, but I think it's at a point where manually editing (particularly round the chord change points) would be doable.

[lmv]

This was the best I could get it;

MP3 using a sampler patch I made of one of my acoustics;

That sounds fairly good, but the MIDI file looks a lot like what I have been obtained with all the converters I tried. It's way too messy. I need something clean that I can work on and change slightly.

Many thanks for your attempt. It shows I am not really doing anything wrong. The technology is just not quite there yet. And who said it would be? Some of us (i.e. me) are spoiled by too much technology and harbor unrealistic expectations.

[donkey tugger]

This was the best I could get it;

MP3 using a sampler patch I made of one of my acoustics;

That sounds fairly good, but the MIDI file looks a lot like what I have been obtained with all the converters I tried. It's way too messy. I need something clean that I can work on and change slightly.

Many thanks for your attempt. It shows I am not really doing anything wrong. The technology is just not quite there yet. And who said it would be? Some of us (i.e. me) are spoiled by too much technology and harbor unrealistic expectations.

Yeah, I think the original is probably just a bit too indistinct and uneven in recording quality for a good conversion (which you can get using the technology with a better starting point). Whilst I'm an ok guitarist for what I do, my fingerpicking is atrocious to abysmal, so I'll let someone else have a go at re-playing it!