ludlows / PESQ

PESQ (Perceptual Evaluation of Speech Quality) Wrapper for Python Users (narrow band and wide band)
https://github.com/ludlows/PESQ
MIT License
518 stars 97 forks source link

Max PESQ #13

Closed MinMolang closed 4 years ago

MinMolang commented 4 years ago

I have tasted with two same audio wave files.(16khz) I got result like this. result

I thought that max PESQ would be 4.5. I'd like to know the python-PESQ max value. I appreciate your pyhon-pesq module. Thanks. XD

ludlows commented 4 years ago

PESQ results principally model mean opinion scores (MOS) that cover a scale from 1 (bad) to 5 (excellent). A mapping function to MOS-LQO is outlined under P.862.1

leo19941227 commented 4 years ago

Standard implementation in pesqmain.c (C code) P.862: Software/P862_annex_A_2005_CD/source/pesqmain.c

if ( err_info->mode == NB_MODE )
{
    err_info->mapped_mos = 0.999f+4.0f/(1.0f+(float)exp((-1.4945f*err_info->pesq_mos+4.6607f)));
}
else
{
    err_info->mapped_mos = 0.999f+4.0f/(1.0f+(float)exp((-1.3669f*err_info->pesq_mos+3.8224f)));
    err_info->pesq_mos = -1.0;
}

Little experiment (correction is highly appreciated if I made any mistake!)

import numpy as np

def mapping(pesq_value, mode='NB'):
    if mode == 'NB':
        return 0.999 + 4.0 / (1.0 + np.exp(-1.4945 * pesq_value + 4.6607))
    else:
        return 0.999 + 4.0 / (1.0 + np.exp(-1.3669 * pesq_value + 3.8224))

nb_min, nb_max = mapping(-0.5, 'NB'), mapping(4.5, 'NB')
wb_min, wb_max = mapping(-0.5, 'WB'), mapping(4.5, 'WB')

print(f'NB range: ({nb_min}, {nb_max})')
print(f'WB range: ({wb_min}, {wb_max})')

Result

NB range: (1.016843313292765, 4.548638319075995)
WB range: (1.0426942267891948, 4.643888749336258)
MinMolang commented 4 years ago

Thank you for let me know NB range and WB range! And I checked P.862.1 in here https://www.itu.int/rec/T-REC-P.862.1-200311-I/en And It said "In addition, the mapping ensures a domain rescaling from –0.5 ... 4.5 to 1.02 ... 4.56. " So now I understand this range is not exactly 1.0 to 4.5. I totally didn't know about P.862 or MOS-LQO these things..! My result is exactly same with @leo19941227 's result. So I will keep in mind that this is P.862.2! Thanks. XD

leo19941227 commented 4 years ago

I am sorry for a small mistake. Actually the code is provided in P.862, however it did follow the mapping function described in P862.1 and P862.2 when it was amended in P862 Amendment 2.

In conclusion, P862.1 proposes a mapping function for narrow band raw pesq value -> MOS-LQO, and P862.2 proposes another one for the wide band case. They are two similar but different mapping functions. Finally, P.862 Amendment 2 implements these two functions in C code.

arunraj-sandeza commented 1 year ago

At the present moment i'm using your code for PESQ, but can someone help me with usage of narrow band and wide band in the code, like why we use nb and wb?

ludlows commented 1 year ago

@arunraj-sandeza thanks for using our code.

regarding your question, the narrow and wide bands are distinguished by the upper frequency bound. the narrow band has a lower frequancy bound 8000Hz. while the wide one has a higher bound 16000Hz. in other words, we could use the sampling rate of your raw audio samples to help you choose the mode.

zdj97 commented 6 months ago

what is the smallest value of pesq?