I hear crackling sound when using 32bit integer numpy arrays to play audio

[physxP]

Code to reproduce the issue:

import numpy as np
import sounddevice as sd

frequency = 1000
sample_rate = 44100
duration  = 2 # seconds
num_frames = int(duration*sample_rate)

t = np.arange(num_frames) / sample_rate
max_amp = 2**31-1 # maximum value of signed int32
signal = max_amp*np.sin(2 * np.pi * frequency * t)
signal = signal.astype(np.int32)
sd.play(signal,samplerate=sample_rate)

Current OS: Windows 10 Things I have tried:

• Running in other high end laptop running windows 10
• Directly using float32 format doesn't cause crackling sound but I need to work with int32 and int24
• Using raw output stream and writing to audio buffer has same effect for int32 and int24
• int16 and int8 play well using same method
• float32 and float16 work well
• int24 has same issue as int32
• Increasing sampling rate from 44.1k to 96k increased the frequency of noise
• Disable enhancements in windows sound driver does nothing
• Uninstalled sound driver - no difference
• Tried giving stereo output with same results
• increasing block size upto 10 times the sample rate
• playing with latency settings
• Checking status of stream - no warning or error

[HaHeho]

What is the reason for that you say that you need to work with int32 (your original data / processing generates that)?

In any case, you should be able use .astype(np.float32) (and potentially scale the amplitude) to make it work, right?

[HaHeho]

I have tested a bit on my system (Win10) with int32:

import numpy as np
import sounddevice as sd

frequency = 1000
sample_rate = 44100
duration = 1  # seconds
amplitude = 0.1  # -20 dB, clean
amplitude = 1  # 0 dB, results in crackles

num_frames = int(duration * sample_rate)
t = np.arange(num_frames) / sample_rate
signal = amplitude * np.sin(2 * np.pi * frequency * t)

print(f'play dtye: {signal.dtype}, max: {np.max(signal):.2f}')
sd.play(signal, samplerate=sample_rate, blocking=True)

signal *= 2 ** 31 - 1  # maximum value of signed int32
signal = signal.astype(np.int32)

print(f'play dtye: {signal.dtype}, max: {np.max(signal):.2f}')
sd.play(signal, samplerate=sample_rate, blocking=True)

What is suspicious is that amplitude = 0.1 gives a clean output with identical volume for float and int. Hence, the amplitude scaling is correct).

However, amplitude = 1 gives the crackling output, as if the signal is clipping, although the value range should be fine, right?

[physxP]

What is the reason for that you say that you need to work with int32 (your original data / processing generates that)?

In any case, you should be able use .astype(np.float32) (and potentially scale the amplitude) to make it work, right?

I am working with a tight constraint on precision. The loss of precision in conversion from int32 to float32 is not desirable. Also i need to work with 24bits and from what I read, I think sound-device only works with int24.I have same issue with int24

[physxP]

I have tested a bit on my system (Win10) with int32:

import numpy as np
import sounddevice as sd

frequency = 1000
sample_rate = 44100
duration = 1  # seconds
amplitude = 0.1  # -20 dB, clean
amplitude = 1  # 0 dB, results in crackles

num_frames = int(duration * sample_rate)
t = np.arange(num_frames) / sample_rate
signal = amplitude * np.sin(2 * np.pi * frequency * t)

print(f'play dtye: {signal.dtype}, max: {np.max(signal):.2f}')
sd.play(signal, samplerate=sample_rate, blocking=True)

signal *= 2 ** 31 - 1  # maximum value of signed int32
signal = signal.astype(np.int32)

print(f'play dtye: {signal.dtype}, max: {np.max(signal):.2f}')
sd.play(signal, samplerate=sample_rate, blocking=True)

What is suspicious is that amplitude = 0.1 gives a clean output with identical volume for float and int. Hence, the amplitude scaling is correct).

However, amplitude = 1 gives the crackling output, as if the signal is clipping, although the value range should be fine, right?

Yes it should work but something wonky is going on. Even if I choose amplitude of 0.99 the crackling sound is gone. Going to check the frequency spectrum of the integer signal to check for noise Update: I checked the code on ubuntu20.04 and it is working without the crackling noise. The problem might be on windows side

[mgeier]

Did you try different host APIs on Windows?

[physxP]

Can you please tell me how can I do that? I couldn't find anything related to changing host APIs in the documentation

[mgeier]

If there are multiple host APIs, your physical devices will appear multiple times in the device list. Each entry contains the host API name after the device name.

So by selecting a specific device from the list, you automatically select a host API at the same time.

I couldn't find anything related to changing host APIs in the documentation

Yeah, I think that's not mentioned explicitly.

Do you have an idea how and where we could document this? Would you like to make a documentation PR for this?

[physxP]

Thanks for the quick response. Changing the host api fixed the issue! I tried all the listed APIs and switching to the "Windows WDM-KS" API fixed that. Lists of APIs and results for future reference:

• MME : Same issue
• Windows DirectSound: Same issue
• ASIO: Same
• Windows WASAPI: Error while playing signal: (PortAudioError: Error opening OutputStream: Invalid device [PaErrorCode -9996])
• Windows WDM-KS: Success!

Do you have an idea how and where we could document this? Would you like to make a documentation PR for this?

Sure, I would love to help. I think we can mention this in Device Selection. We can modify the heading to something like 'Device and Host API Selection'. We can also add a troubleshooting section in which we can mention common issues and their fixes starting with this one. What do you think?

[mgeier]

Sounds good, I'm looking forward to a PR!

About the crackling issue: It's great to hear that it works with WDM-KS. I guess the issues with the other host APIs will have to be fixed in https://github.com/PortAudio/portaudio. Would you like to open an issue there?