Advanced beat swapping powered by madmom.
For most people I recommend using Hugging Face or Google Colab. However if you run it locally, you will have access to more advanced features that I haven't added to Hugging Face yet, like using samples, mixing multiple songs, presets.
First I recommend creating a new environment to avoid dependency issues. With conda, you can do that by running conda create --name beat_manipulator.
Then run those commands depending on what python you have:
conda install pip cython mido numpy scipy pysoundfile librosa ffmpeg-python pytest pyaudio pyfftw
pip install madmom pedalboard
conda install pip cython mido numpy scipy pysoundfile librosa ffmpeg-python pytest pyaudio pyfftw
pip install pedalboard
pip install git+https://github.com/CPJKU/madmom
pip install numpy cython soundfile ffmpeg-python pedalboard librosa
pip install madmom
pip install numpy cython soundfile ffmpeg-python pedalboard librosa
pip install git+https://github.com/CPJKU/madmom
After installing all necessary libraries, to download beat manipulator, download and extract this repo using green "Code" button > Download ZIP, or run git clone https://github.com/stunlocked1/beat_manipulator. You can now open examples.py, jupiter.ipynb, or app.py for gradio interface.
First, import beat_manipulator and load a song
import beat_manipulator as bm
your_song = bm.song(audio = 'path or numpy array')
It accepts absolute or relative path to audio file, or you can directly load audio array into it. If you are loading audio array directly, make sure to also add sr=sample_rate argument. Array should be in -1 to 1 format, which is how most libraries load audio.
Now, generate the beatmap:
your_song.beatmap_generate()
Beatmap is generated using madmom library. When you generate it for the first time, it might take up to a minute. However all beatmaps are saved to beat_manipulator/beatmaps, so when you load the same file for the second time, it will be instant.
You can access beatmap in your_song.beatmap variable. It is a list of values that represent position of each beat in samples.
After generating the beatmap, you can do a bunch of stuff.
Song object supports slicing.
your_song[5]will return audio of the 5th beat, indexing starts from 1.your_song[4:8.5]returns audio starting from 4th beat, ending halfway between 8th and 9th beat.your_song[0:1]is equivalent toyour_song[1]your_song['your_pattern']returns a beatswapped audio using the pattern you provided. Beatswapping with patterns is the main feature if this app and you can do a whole bunch of stuff with them. There is a section below that explains how to write those patterns.
Another way to beatswap is:
your_song.beatswap(pattern = '1, 3, 2, 4', scale = 1, shift = 0, length = None)
This one doesn't return anything, instead it modifies the song in place.
You can also beatwap and write audio in one line:
bm.beatswap(song = 'path or numpy array', pattern = '1, 3, 2, 4', scale = 1, shift = 0, output = '')
scale = 0.5 will insert a new beat position between every existing beat position in the beatmap. That allows you to make patterns on smaller intervals.
scale = 2, on the other hand, will merge every two beat positions in the beatmap. Useful, for example, when beat map detection puts sees BPM as two times faster than it actually is, and puts beats in between every actual beat.
To scale the beatmap, you can use your_song.beatmap_scale(0.5), or specify scale directly in your_song.beatswap(..., scale = float)
Shifts the beatmap, in beats. For example, if you want to remove 4th beat every four beats, you can do it by writing 1, 2, 3, 4!.
However sometimes it doesn't properly detect which beat is first, and for example remove 2nd beat every 4 beats instead. In that case, if you want 4th beat, use shift = 2. Also sometimes beats are detected right in between actual beats, so shift = 0.5 or -0.5 will fix it.
To shift the beatmap, you can use your_song.beatmap_shift(0.5), or specify shift directly in your_song.beatswap(..., shift = float)
When you specify shift in a beatswap function, it applies before scale for consistency.
If you run your_song.beatmap_save_settings(scale: float, shift: float), it will save a file in beat_manipulator/beatmaps with your scale and shift. That way, next time you load that song, it will automatically apply those scale and shift values.
To write audio, use my_song.write(output = ''). If output is empty string, this will write the song next to your .py file, using the original filename.
The pattern syntax is quite powerful and you can do a whole bunch of stuff with it. Basic syntax is - 1, 3, 2, 4 means every 4 beats, swap 2nd and 3rd beats, but you can do much more, like applying audio effects, shuffling beats, slicing them, mixing two songs, adding samples, sidechain.
You can use spaces freely in patterns for formatting. Most other symbols have some use though. Here is how to write patterns:
1- 1st beat;1>0.5- first half of first beat- `1<0.5 - second half of first beat
0:0.5- range of beats, this also means first half of first beat, but with this you can do complex stuff like1.25:1.5. However this one is a bit more confusing because indexing starts from 0, so 1:2 is second beat, not first.- Also sometimes it is more convenient to use smaller
scale, like 0.5 or 0.25, instead of slicing beats.
1, 3, 2, 4means 3rd and 2nds beats will be swapped every 4 beats. Happens every 4 beats because 4 is the biggest number in the pattern.1, 2, 3, 4!means every 4 beats, it will remove the 4th beat.!allows you to skip a beat but it still counts for pattern size.- Specifying pattern length:
pattern = '1,2,3', length = 4is another way to remove 4th beat every 4 beats. 1,4skips 2nd and 3rd beat1; 2plays 1st and 2nd beat same time. They will automatically be normalzied to avoid clipping. Length of the first beat will be preserved.
, and ; are beat joining operators, that join beats together. Here are all available joiners:
,puts the beat next to previous beat;puts the beat on top of previous beat. Normalizes the volume to avoid clipping. If previous beat is shorter, your beat will be shortened to match it.^multiplies previous beat by your beat. This can be used for fake sidechain.$adds the beat on top of previous beat + sidechains previous beat by your beat.
beats can be followed by effects. For example 1s0.75 means take first beat and play it at 0.75x speed. Here are all available effects:
s- speed.1s2means first beat will be played at 2x speed.r- reverse.1rmeans first beat will have reversed audio.v- volume.1v0.5means 1st beat will have 50% volumed- downsample, or 8-bit sound.1d10will downsample the first beat so that it sounds 8-bit. Good values start above 7.b- bitcrush.1b4will bitcrush it.g- gradient, sounds like highpass.1g1is the recommended valuec- channel. If not followed by number, swaps channels. If followed by 0, plays only left channel. If 1, only right channel- mixing effects -
1s2rd8- take first beat, play at 2x speed, reversed, and downsampled. You can also define your own effects. CheckBM_EFFECTSdictionary frombeat_manipulator/effects.py, this is where it reads effects from. You can add new stuff to that dictionary and use your own effects this way, or specify your own dictionary when usingyour_song.beatswap(..., effects: dict)with youreffectsargument. By default that argument points toBM_EFFECTSdictionary.
mathematical expressions with +, -, *, /, and ** are supported. For example, if you write 1/3 anywhere in the pattern, to slice beats or as effect value, it will be replaced by 0.3333...
To use samples, provide them in samples argument to your_song.beatswap(..., samples: dict). The dictionary should look like this:
{
'sample name' : 'path to your sample or numpy array of your sample or bm.song object',
'sample name 2' : 'path or audio 2',
...
}
It supports both loading audio files from a path, and directly loading arrays.
Then in pattern, you can use quotes (', ", or `) to access samples. For example: 1; "sample_name" will put that sample on top of 1st beat. Samples are treated just like beats, you can apply effects to them, use any joining operators.
You can also slice samples: "sample_name">0.5 means first half of the sample.
You can use list with samples instead of a dictionary. In that case, you can access them by their index in the list, for example "1" is the 1st sample.
Add the second song that you want to mix to the samples argument dictionary or list, as described above. It can load path to a file, directly load a numpy array, or a bm.song object.
The difference is, instead of using quotes, for songs you use square brackets: [song_name]
[song_name]4 means fourth beat of that song. So you can do stuff like 1, [song_name]2, which will alternate beats between your two songs.
iwill be replaced by current position, e.g.i, i, i, i+1is equvalent to1, 2, 3, 4 + 1, or1, 2, 3, 5.#will add shuffle all beats with the same number after it.1#1, 2#2, 3#1, 4#2, 5#1, 6#2, 7#1, 8#2will shuffle 1st, 3rd, 5th and 7th beats (the are in 1st group), and 2nd, 4th, 6th and 8th beats - from 2nd shuffle group.!skips that beat. If you want to remove every 4th beat, you can't just do1, 2, 3, because that would simply play every 3 beats. So to play 3 beats every 4 beats, you can write1, 2, 3, 4!?makes that beat not count for pattern size. For example,1, 2, 3, 8will normally repeat every 8 beats because 8 is the highest number, but1, 2, 3, 8?will repeat every 3 beats.@allows you to take a random beat with the following syntax: @start_stop_step. For example,@1_4_0.5means it will take a random beat out of 1st, 1.5, 2nd, 2.5, 3rd, 3.5, and 4th. It will take whole beat, so you can also add>0.5to take only first half.%- for very advanced patterns you can create variables from various metrics. For example,%vwill create a variable with average volume of that beat, and all following%will be replaced by that variable until you create a new one. Useful for applying different effects based on different song metrics. All metrics are inbeat_manipulator/metrics.py.
You can write special commands into the pattern argument instead of actual patterns.
reverse- plays all beats in reverse chronological ordershuffle- shuffles all beatstest- puts different pitched cowbells on each beat, useful for testing beat detection and adjusting it using scale and shift. Each cowbell is 1 beat, highest pitched cowbell is the 1st beat, lowest pitched - 4th.
You should be able to use all of the above operators in any combination, as complex as you want. Very low scales should also be fine, up to 0.001.
You can create cool images based on beat positions. Each song produces its own unique image. Write:
your_song.image_generate()
image will be saved as a numpy array to your_song.image variable. To export it to a file, use:
your_song.image_write()
The image will by default be resized to 4096x4096. It is also possible to export original image, which usually is too big for most image viewers to handle it. However the cool thing is that you can apply image effects to it, and then turn it back into audio. I will soon add info on how to do that.
bm.beatswap(song = 'path or numpy array', pattern = '1,3,2,4', scale=1, shift=0, output='')
allows you to beatswap and write a song loaded from path or numpy array in one line. Returns path to the exported beatswapped song file.
bm.image(song = 'path or numpy array', max_size = 4096, scale=1, shift=0, output='')
creates an image and writes it in one line, returns path to exported image.
bm.osu.generate(song='path or numpy array', difficulties = [0.2, 0.1, 0.05, 0.025, 0.01, 0.0075, 0.005, 0.0025, 0.0001])
generates an osu! beatmap (uses madmom beat processor and peak detection). Writes an .osz file that you can install by opening it with osu! and returns path to it.
there are some patterns in beat_manipulator/presets.yaml file. Those are supposed to be used on normalized beat maps, where kick + snare is two beats, so make sure to adjust beatmaps using scale and shift.
To use one of the presets from that file, write:
bm.presets.use(song = song, preset = 'preset name', scale = 1, shift = 0)
I will clean up the code and then it will be possible to actually understand what is going on. That will happen... At some point