Hello,
I have been working on the decorrelation of audio signals while following this paper :
"Article The Decorrelation of Audio Signals and Its Impact on Spatial Imagery
"So far I have generated impulse responses in such a way that magnitude is 1 in all frequency spectrum, while random phase generation between -π to π. I correlated these signals and I am getting the required correlation.
Now moving on, when I convolve this impulse response with my audio signal, results in the case of decorrelation are good but the audio becomes noisy and distorted. I am attaching the reference wav file and and decorrelated wav file.
Some points regarding computation:
- All audio data computations are in float range [-1.0:1.0].
- I am using Python to perform all the calculations
- The input reference file is in Short (PCM 16 bit) format which is converted to float range [-1.0:1.0].
- I have saved convolved_decorrelated_output.wav in both formats PCM 16 bit short and Float as mentioned above. Still getting the same results. So there are no issues regarding data type conversion as per my knowledge.
Any help will be appreciated.
Regards,
Khubaib Ahmad
Dear Khubaib Ahmad ,
It is clear that the generated function for the decorrelation of the audio file caused this distortion.
I would like to give you a conceptual answer at first:
The decorrelation function must randomize the phase and keeps the amplitude constant. It must also have the same frequency range of the audio signal.
So, I would propose that you plot the frequency response of the audio signal and the decorrelation function and show them to us to see the cause of the distortion.
Best wishes
Abdelhalim abdelnaby Zekry Sir, Please find attached frequency response of my decorrelation signal that I generated by following the paper.
I am also attaching the convoluted generated output from this response.
Dear Khubaib Ahmad ,
I am sorry.
The figures are not clear.
After my opinion, the frequency response of the signal after decorrelation must be
the same at that before the decorrelation.
The decorrelation must randomize the phase of the signal.
Randomizing the phase of the signal must not affect the sound as the ear is not sensitive to the phase variation. My answer is based on this fact.
So the decorrelation function must have uniform flat frequency response.
Best wishes
Sir Abdelhalim abdelnaby Zekry , the technique so far I followed through the paper was that I generated Impulse response of length 1024 in the frequency domain with Unit Magnitude and Random phase variation in range [-π,π].
Then I took ifft of the signal and got an impulse response in the time domain of the same length i.e 1024.
Now, when I convoluted this with my audio signal by making chunks of my audio signal of the same length i.e 1024, It caused distortion and robotic voice. I am using the convolution of python's numpy with the mode "same".
Dear Khubaib Ahmad ,
welcome again!
Fine!
I expect the distortion is due to dividing the long sound message into partitions.
One has to grantee the matching of the signals at the boundaries. Or one has increase the length of the partitions at least to get out t5he cause of the distortion.
Best wishes
Sir Abdelhalim abdelnaby Zekry ,
I performed an experiment of splitting the data into chunks of 21ms (1024 shorts) at a sampling rate of 48KHz. I didn't perform any signal processing on the chunks and recombined them.
There was no issue in the sound regarding dividing the sound into partitions.
Dear Khubaib Ahmad ,
The partitioning itself without performing any signal processing does not lead to any distortion. But performing signal processing the windowing may lead to distortion. So, please continue the experimentation to get the cause of the distortion. As i said to you you must study the effect of the chunk length on the signal distortion.
I would propose that you make the convolution in frequency domain and then making the ifft.
Best wishes
Sir Abdelhalim abdelnaby Zekry . I have found the issue. It is with the chunks creation and performing convolution. I tried the following two tests:
- I convolved my impulse response of 1024(21ms) with a complete audio signal of 6 sec. I got the required decorrelation and the signal wasn't distorted or had a robotic voice.
- Then I convolved by making chunks of 21ms of my audio. Results became distorted.
So, Yes the culprit is signal processing (convolution) of the impulse response of length 1024 with the same length of data.
Now, related to your proposal Sir, I have performed convolution in the frequency domain and took ifft. Still, I am getting the same results (in the case of chunks)
P.S: Data is in the float with a sampling rate 48KHz.
Some useful information: 21ms of data = 1024 length of a float array
Dear Khubaib Ahmad
great !
It is time to mark my answer as best answer!
Thank you!
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Programming Languages
- Python