I got a question, how exactly did you produce the files in the http://www.repository.voxforge1.org/downloads/SpeechCorpus/Trunk/Audio/Main/ directory ?
I know that they are somehow produced from the "Original" secion of the downloads because the files have the same names as the files in the "Original" section and the content appears to sound exactly the same. I immediately identified that the files are downsampled but I could not figure out how they were downsampled (appears like some elaborate algorithm was used to do the downsampling).
I am trying to build some sort of automatic downsampling tester (the script takes the original data, downsamples them using sox to the samplerate I tell it to use, and then runs training of the model on these downsampled data). I wanted to use the data from http://www.repository.voxforge1.org/downloads/SpeechCorpus/Trunk/Audio/Main/ as test data of sorts (tell the script that the target samplerate is 8/16 kHz, let it generate the data and then compare the generated data to these files) but I realized that I am unable to reproduce the process used to produce these data.
I suspected that you used "sox" to derive these files from their original files, so I installed sox 14.4.1 and tried to use "sox b0284.wav test.wav rate 16000" but the resulting file differs from that of yours way too much. I was thinking that the dithering is the cause but then I did some experimenting and I realized that the dithering introduces only very subtle differences into the file, which are not consistent with the differences that are between my file and your file. Reading the sox documentation revealed that the sox resampling algorithm is "very configurable" and there were about 2 options which accept over 30 different settings and the response of the algorithm to the changes of these options is wildly different so there is no way to intelligently search for the "correct" values of these options.
So I need to know whether you are using "sox" and what version you are using and what options you use when downsampling the files.
Sory for the delay, be travelling for the past couple of weeks, back in the middle of August...
>I suspected that you used "sox" to derive these files from their original files
Thank you for the pointer, it is now clear for me that you just used sox with its defaults to do the downsampling.
But I still have unanswered question about what version you are using, certainly it is not SOX v14.4.1 as the output from it is way too much different and I am not able to see whether it is better or worse than that of yours.
You know, I am trying to construct some sort of audio preprocessing and tried to use your files as testcases (whether I set everything up correctly etc) but right now the results are that output from my process differs too much from the output of your process and I am unable to figure out where the difference is and what significance it has.
I would suggest adding '-D' to the code so that it produces exactly the same file each time the code is run and to record somewhere (in the comment?) what version of sox is being used. The idea would be that if for example I want to use the scripts, I can run them over a couple of your input files and compare the results to your output files and see if there is a match (meaning that I have it set up correctly) or not (an error was made somewhere at my side).
I think I found a way to figure this request out.
There is a file called installation.txt which describes how to install the scripts. Based on the content I figured out that the OS used to run the scripts is Fedora Linux. The sox is simply pulled out of the Fedora's repository according to the file (verified also by looking at the oldest revision available and it still says so) so all I need to do is lookup which version of sox Fedora had at the time approximately when the scripts were installed on your working site (about 3 years in the past as of today).
So far so good. I am going to hunt the thing down and then try to install the scripts and play with them until I get results I am satisfied with.