Description

I am new to this project and have been searching for time-stretching and pitch-shifting functionalities in python when I found this so, to test its working, I ran the "Example Usage" code from the pyrubberband github with one of my file but I got this error (image below).

I tried finding the solution to this on web, but had no leads as of now.

Versions

Windows-10-10.0.17134-SP0
Python 3.7.3 (default, Apr 24 2019, 15:29:51) [MSC v.1915 64 bit (AMD64)]
NumPy 1.16.4
SoundFile 0.9.0

Moreover

I think the problem is with the OS (I'm using windows, but found OS X codes around the web), and by the way I also tried pip install rubberband-cli in the prompt, cuz the error said "rubberband-cli not installed" but that generated another error :-

Collecting rubberband-cli
ERROR: Could not find a version that satisfies the requirement rubberband-cli (from versions: none)
ERROR: No matching distribution found for rubberband-cli

Later, I downloaded rubberband cli from here and tried installing that as well but that too didn't help.

So finally I am here, hoping to get some positive response from the community. Thanks !

I added a function that allows users to perform the timemap stretch option of rubberband (see doc here). It is largely based on the current code.

I've also added a test in the tests file.

Here is an example of time map that stretches a 5 seconds audio file by: Leaving second 0 to 1 as is Speeding up seconds 1 to 3 by a factor 2. (it corresponds to seconds 1 to 2 in the stretched audio) Leaving the rest of the audio as is.

time_map = [(0, 0), (sr, sr), (3*sr, 2*sr), (5*sr, 4*sr)]

I really like librosa but importing it only for the read/write capability can make things complicated. There are other libraries which only focus on audio I/O like pysoundfile and therefore have less dependencies and are more flexible.

This PR therefore removes librosa dependency and replace it with pysoundfile. I know that pysoundfile is not optimal because it not pure python but since timestretching/pitch shifting is often applied on general audio files a library which supports a large number of input and output formats like portaudio is beneficial. Also to my knowledge there are currently no better solutions for python.

I've also added a note howto easily install rubberband cli on mac by providing a homebrew recipe

pysoundfile is deprecated: https://github.com/bastibe/PySoundFile

What does this implement/fix? Explain your changes.

The PySoundFile page on GitHub indicates that PySoundFile was renamed to SoundFile.

The install command is now pip install soundfile.

Any other comments?

I had to comment out the pytest lines in setup.cfg in order to run the tests. But, the tests did pass!

Running pytest with setup.cfg as is results in:

↪ pytest
================================================================================ test session starts ================================================================================
platform darwin -- Python 3.6.12, pytest-6.1.2, py-1.9.0, pluggy-0.13.1
rootdir: /Users/nw_henderson/github.com/nwh/pyrubberband, configfile: setup.cfg
plugins: cov-2.10.1, pep8-1.0.6
collected 0 items / 1 error                                                                                                                                                         
Coverage.py warning: No data was collected. (no-data-collected)

====================================================================================== ERRORS =======================================================================================
___________________________________________________________________________ ERROR collecting test session ___________________________________________________________________________
Direct construction of Pep8Item has been deprecated, please use Pep8Item.from_parent.
See https://docs.pytest.org/en/stable/deprecations.html#node-construction-changed-to-node-from-parent for more details.

---------- coverage: platform darwin, python 3.6.12-final-0 ----------
Name                       Stmts   Miss  Cover   Missing
--------------------------------------------------------
pyrubberband/__init__.py       3      3     0%   4-7
pyrubberband/pyrb.py          72     72     0%   3-257
pyrubberband/version.py        1      1     0%   4
--------------------------------------------------------
TOTAL                         76     76     0%

============================================================================== short test summary info ==============================================================================
ERROR 
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! Interrupted: 1 error during collection !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
================================================================================= 1 error in 0.09s ==================================================================================

Reference Issue

N/A

What does this implement/fix? Explain your changes.

This adds a function ableton_asd_to_time_map for parsing Ableton .asd warp files into a time_map list to be used with pyrb.timemap_stretch.

Any other comments?

Example usage:

import pyrb
import librosa

sr = 44100
bpm = 120.

y = librosa.load('song.wav', sr=sr)[0]
time_map = pyrb.ableton_asd_to_time_map('song.wav.asd', y, sr, bpm)
new_song = pyrb.timemap_stretch(y, sr, time_map)

from scipy import io
import scipy.io.wavfile
io.wavfile.write('new_song.wav', sr, new_song )

I wrote the code after reading https://github.com/jtxx000/extract-warp-markers. That project showed how to extract the BPM. I did some extra probing to find out how to get other fields such as loop_start, loop_end, start_marker, end_marker, loop_on. I think having that information could be really useful for other developers. Perhaps pyrubberband could handle looping one day.

It's possible that this function could break if Ableton changes their ASD format... But I still think it's worth sharing. I'm using Ableton 10.1.30.

sf.read defaults to reading files as float64, which causes pyrubberband to implicitly convert all processed audio into float64 regardless of the original data type. This change makes pyrubberband preserve the data type of the input by reading it back in the same format.

I followed this example that I found on Medium but it doesn't seem to work. When the playbackRate is increased, how can I correct the pitch so that it is the same as the original audio?

I have also tried the following by passing the output of time_stretch to the input of pitch_shift

# Play back at 1.5X speed
y_stretch = pyrb.time_stretch(audioData, sampleRate, audioSpeed)
# Play back two 1.5x tones
y_shift = pyrb.pitch_shift(y_stretch, sampleRate, audioSpeed)
soundfile.write(f'{tmp_name}.wav', y_shift, sampleRate, format='wav')

Reference Issue

Fixes Issue #21

What does this implement/fix? Explain your changes.

Arguments to CLI, parsed through rbargs can be single or dual valued, but currently breaks if it's single valued. for e.g. pyrb.pitch_shift(wav_data, n_semitones, rbargs={'--formant'}) won't work.

This current implementation allows us to restore that functionality by parsing an empty value '' to the CLI.

This fix allows us to parse flags through in the form of rbargs={'--flag' : ''} while retaining the kwargs format.

Any other comments?

I do feel that this leaves us a bit wanting for something nicer but it's just a step forward to allow some functionality for now. We can explore making the extra argument as None instead of '', but I figured this is probably just as intuitive for these occasional use-cases.

Description

Some flags in rubberband are not followed by a value, e.g. --no-lamination or --pitch-hq. The current implementation always expects rbargs as pairs of keyword and argument (in dict format), making it impossible to use flags that don't take values.

Steps/Code to Reproduce

import pyrubberband

# Generate a random signal and time-stretch it
sr = 22050
y = np.random.randn(5 * sr)

y_stretch = pyrubberband.time_stretch(y, sr, rate=1.5, rbargs={"-c": "6"})  # works
y_stretch = pyrubberband.time_stretch(y, sr, rate=1.5, rbargs={"--pitch-hq": ""})  # fails
y_stretch = pyrubberband.time_stretch(y, sr, rate=1.5, rbargs={"--pitch-hq": " "})  # fails
y_stretch = pyrubberband.time_stretch(y, sr, rate=1.5, rbargs={"--pitch-hq": None})  # fails

Expected Results

Should be possible to call no-value flags such as --pitch-hq. Included some examples above of syntax that could be supported with dicts (e.g. None). Alternative dict could be replaced with e.g. tuple that can have either 1 or 2 elements.

Actual Results

---------------------------------------------------------------------------
CalledProcessError                        Traceback (most recent call last)
<ipython-input-45-8758af91c0ad> in <module>
      5 filename = '/Users/salamon/dev/scaper/tests/data/audio/foreground/human_voice/42-Human-Vocal-Voice-all-aboard_edit.wav'
      6 audio, sr = soundfile.read(filename)
----> 7 audio_pitch = pyrubberband.pitch_shift(audio, sr, 1, rbargs=params)
      8 Audio(data=audio_pitch.T, rate=sr)

~/dev/miniconda3/envs/scaper35/lib/python3.5/site-packages/pyrubberband/pyrb.py in pitch_shift(y, sr, n_steps, rbargs)
    255     rbargs.setdefault('--pitch', n_steps)
    256 
--> 257     return __rubberband(y, sr, **rbargs)

~/dev/miniconda3/envs/scaper35/lib/python3.5/site-packages/pyrubberband/pyrb.py in __rubberband(y, sr, **kwargs)
     72         arguments.extend([infile, outfile])
     73 
---> 74         subprocess.check_call(arguments, stdout=DEVNULL, stderr=DEVNULL)
     75 
     76         # Load the processed audio.

~/dev/miniconda3/envs/scaper35/lib/python3.5/subprocess.py in check_call(*popenargs, **kwargs)
    269         if cmd is None:
    270             cmd = popenargs[0]
--> 271         raise CalledProcessError(retcode, cmd)
    272     return 0
    273 

CalledProcessError: Command '['rubberband', '-q', '-c', '6', '--pitch-hq', ' ', '--pitch', '1', '/var/folders/j4/2cgkdym179b76zljmlm1d6yc0000gn/T/tmpu311s4a2.wav', '/var/folders/j4/2cgkdym179b76zljmlm1d6yc0000gn/T/tmpermm7whf.wav']' returned non-zero exit status 2

Versions

Darwin-18.2.0-x86_64-i386-64bit Python 3.5.6 |Anaconda, Inc.| (default, Aug 26 2018, 16:30:03) [GCC 4.2.1 Compatible Clang 4.0.1 (tags/RELEASE_401/final)] NumPy 1.18.2 SoundFile 0.9.0

Reference Issue

No reference issue

What does this implement/fix? Explain your changes.

Implement a function that were not wrapped: frequency multiply: To multiply all frequencies by a certain factor, helpful with some stuff that shifting by semitones cannot achieve. Is the -f option in the CLI

tempo shift: Instead of calculating manually the ratio conversion for the tempo, use the function already included in rubberband

Pyrubberband will hang when faced with low-sr signals unless the --window-long flag is set. This PR adds that flag for signals below 210Hz and allows the process to complete.

I've noticed that when you try to use rubberband on very low-samplerate signals (anything 210Hz and below) rubberband will hang.

These both hang:

y = np.random.rand(1024)

shifted = pyrubberband.pitch_shift(y=y, sr=210, n_steps=2)
stretched = pyrubberband.time_stretch(y=y, sr=210, rate=0.5)

These both work

y = np.random.rand(1024)

shifted = pyrubberband.pitch_shift(y=y, sr=211, n_steps=2)
stretched = pyrubberband.time_stretch(y=y, sr=211, rate=0.5)

The rubberband CLI has the same behavior; it hangs during the first studying pass at 99% CPU until you send it a break:

$ rubberband -d3 -t 2.0 /tmp/tmpzGl2Zt.wav /tmp/tmpzGl2Zt_2.wav
Using crispness level: 5 (Crisp monophonic instrumental)
Using time ratio 2 and frequency ratio 1
RubberBandStretcher::Impl::Impl: rate = 125, options = 16
configure: effective ratio = 2
configure: analysis window size = 8, synthesis window size = 8, fft size = 8, increment = 0 (approx output increment = 0)
configure: outbuf size = 8192
Window area: 0.5; synthesis window area: 0.5
FFT::FFT(8): using implementation: fftw
Not real time mode: prefilling
configure: effective ratio = 2
configure: analysis window size = 8, synthesis window size = 8, fft size = 8, increment = 0 (approx output increment = 0)
configure: outbuf size = 8192
Not real time mode: prefilling
configure: effective ratio = 2
configure: analysis window size = 8, synthesis window size = 8, fft size = 8, increment = 0 (approx output increment = 0)
configure: outbuf size = 8192
Pass 1: Studying...

Interestingly, this only seems to be an issue when rate for time_stretch < 1 or when n_steps for pitch_shift > 0, otherwise it runs fine.

I haven't had time to dig into the rubberband source and likely will not any time soon but I've noticed that passing the --window-long flag will allow it to run to completion. IE these both run fine:

y = np.random.rand(1024)

shifted = pyrubberband.pitch_shift(y=y, sr=210, n_steps=2, rbargs={"-q":"--window-long"})
stretched = pyrubberband.time_stretch(y=y, sr=210, rate=0.5, rbargs={"-q":"--window-long"})

It's hacky, but would it make sense to simply add a conditional that would add the --window-long flag to any sr <= 210?