RNNoise filter seems to cause a voice distortion

[MGG1996]

OS and Version (please complete the following information):

• Operating System: Windows 11 23H2
• Shotcut Version: [e.g. 26.6.25]

Describe the bug
I add a RNNoise filter and the voice performance was intact during the preview, but after actual export, there was significant distortion in the voice。

Expected behavior
The audio should be the same as it was previewed when there is a RNNoise filter.

[ddennedy]

Unable to reproduce. Maybe you are overly compressing the audio with the choice of audio codec and bitrate or quality level. Here is the input file I used for test

background_noise.mp3

And the output with black video

rnnoise.mp4

[bmatherly]

@MGG1996 can you provide the encoding parameters for your export file? In particular, I wonder about the sampling frequency.

[MGG1996]

Thanks for all your reply.

I tested with the background_noise.mp3 above, and it didn't happen. Again, I extracted a short segment from my original video and was unable to reproduce the issue.

I guess this might be related to my other operations. In my raw materials, due to issues with the recording equipment, there were fluctuations in the volume of the voice before and after, especially in the later stages. So I added a volume gain filter to it and set some keyframes. A portion of the gain reached 24dB, so the background noise was also amplified (but only exists in this segment).

To eliminate the background noise, I added an RNNoise filter and set the suppression rate to 70%. At this moment, during the preview process, the background noise can be eliminated perfectly, and the vocals are normal and perfect.

However, after exporting the video, there was a noticeable distortion in the human voice, which no longer belonged to a normal person's voice. So what's strange about me is that even though there were no issues during the preview process, the exported result doesn't match the preview.

I originally intended to reproduce this issue by capturing a segment from the original video, as the original video lasted for an hour and a half and was as large as 5G, making it difficult to upload. Unfortunately, I was unable to successfully reproduce the issue using the 10 second clip I captured.

Could it be the length of the video that caused this issue? Or is it because I updated Shotcut to version 26.6.25 yesterday but didn't restart my computer? I exported it twice last night and the result was the same.

My encoding parameters are:

[bmatherly]

Thanks for the information. When I export with 48khz, it sounds good. When I export with 44.1khz, the sound is slower and lower pitch. Is that what it sounds like for you, too? Can you try a test to export with 48khz and see if it sounds good for you?

[MGG1996]

Thanks for the information. When I export with 48khz, it sounds good. When I export with 44.1khz, the sound is slower and lower pitch. Is that what it sounds like for you, too? Can you try a test to export with 48khz and see if it sounds good for you?

Thanks for your efforts. I had exported with both 48khz and 44.1khz yesterday, but the result is all the same. At that time, the actual listening experience was not just slightly flawed, but clearly had completely transformed into a voice that did not resemble that of an ordinary person, so the problem should not be here.

I am now a little skeptical whether it is because I did not restart after updating shotcut yesterday, because the 10 second video clip I just captured did not reproduce yesterday's problem.

I am currently exporting a full hour and a half worth of audio to see if it is related to the video duration and the volume gain filter I added.

If the exported long-term audio still has the same problem, I will try to package and upload only the audio project. Fortunately, this audio is much smaller than the video.

By the way, does this export process depend on the ffmpeg version installed on my local machine. Coincidentally, I also updated the ffmpeg version of my device from the stable version to the latest master version yesterday, and found that there was an issue with embedding chapter information into the video using ffmpeg. However, everything was normal before updating ffmpeg.

[bmatherly]

Thanks for the response. So, in your testing, 44.1 export sounds the same as 48? Because they sound different in my testing. I will investigate that.

[MGG1996]

Thanks for the response. So, in your testing, 44.1 export sounds the same as 48? Because they sound different in my testing. I will investigate that.

At least based on my original files, the results exported at the two sampling rates cannot be clearly perceived by the human ear.

The default audio sampling rate for Shotcut is 48kHz, but the sampling rate of my audio file itself is 44.1kHz. Therefore, when exporting for the first time, it was 48kHz. After checking some information, the second export kept the sampling rate consistent with my original material. Unfortunately, there was no significant difference in the results.

[MGG1996]

I export only the audio again, unfortunatelly, the problem is still here.

While, the project size reaches 288M, exceeding the upload limit. @bmatherly Could you please leave an email address, I can't think of a better way to send the project to you.

Encode settings:

[bmatherly]

Before you send me any project, let me see if I can get 44100Hz export working for myself and then you can test a build with your project.

[MGG1996]

Before you send me any project, let me see if I can get 44100Hz export working for myself and then you can test a build with your project.

Thanks for your reponse. During your attempt, I suggest:

1. Use a sufficiently long audio, preferably no less than 1 hour
2. Add a long-term volume gain filter, set keyframes with fluctuating gain values, and set larger gain values on certain keyframes, such as 24dB

This may make it easier to trigger the problem I encountered.

Thank you for your efforts.