forked from spxnn/eaqual
-
Notifications
You must be signed in to change notification settings - Fork 2
ewan-xu/eaqual
Folders and files
Name | Name | Last commit message | Last commit date | |
---|---|---|---|---|
Repository files navigation
Contents 1. What is EAQUAL? 2. How to use EAQUAL 3. Description of Model Output Variables (MOVs) 4. Description of Error Messages 5. The algorithm 6. License of EAQUAL 7. HTML documentation of source code 1. What is EAQUAL? The intention of EAQUAL is to provide a objective quality measure for coded/ decoded audio files. Currently the code is based on the ITU-R recommendation BS.1387. EAQUAL will not supersede listening tests, but can be a useful tool to support listening tests, categorize different coding algorithms and find bugs in algorithms. The more input files are taken for the analysis, the better the results of EAQUAL will fit the reality. 2. How to use EAQUAL To use EAQUAL, you have to provide the reference file, which is the original PCM data (prefereably 16bit, 48kHz), and the test file, which is the coded and decoded audio file and has the *same* audio format. Invalid file formats are samplerates < 44100Hz or more than two channels. The most interesting output value of EAQUAL is the ODG (Objective Difference Grade). An ODG of -4 means a very annoying disturbance, while an ODG of 0 means that there is no perceptible difference. Usage: EAQUAL [-options] -fref reference_file_path -ftest test_file_path Options: -bits specify number of bits in audio file (only needed with raw input, default: -bits 16) -blockout get intermediate results for blocks: -blockout value file_path (valid values are: All, ODG, DI, BWRef, BWTest, NMR, ADB, EHS, ModDiff1, ModDiff2, NoiseLoud, DetProb, RDF, NMRSpec, ModDiffSpec, DetProbSpec, NLSpec). The values will be plotted in the text file file_path for debugging purposes -ch specify number of channels in audio file (only needed with raw input, default: -ch 2) -forget enable forget factor (only applied for output value MFPD: rate distortions at beginning of audio files as less important) -fref reference (original) audio file -ftest audio file under test -h print help -offset shift test buffer for offset samples (that means, discard those samples). If offset <0 the reference buffer is shifted (default: -offset 0) -silent no output to stdio, but in an ASCII file specified after this switch -srate specify sample rate of audio file (only needed with raw input, default: -srate 48000) -spl specify listening level in dBSPL (default: -spl 92) -ver output program version Useful tips: - Both input files have to be time aligned. Many decoders add a frame of silence to the beginning of the decoded file. If this is the case, EAQUAL will not work correctly. Please use the ?offset switch to adjust the delay between reference and test file ? if offset is positive then the test file is delayed and vice versa. - EAQUAL will not work as intended if the file length (of both, reference and test signal) is not sufficient. To achieve reasonable results, a minimal file length of 10-20s is strongly recommended, the longer the better. - The recommendation ITU-R BS.1387 on which EAQUAL is based was developed for high quality codecs with little difference between reference and test signal. If you are trying to use EAQUAL to rank e.g. speech codecs it will fail. If you must use EAQUAL for middle quality codecs then use the DI-output value as ranking, not the ODG (Objective Difference Grade), which is only a shifted and nonlinear (sigmoid) transformed version of the DI. 3. Description of Model Output Variables (MOVs) ODG: Objective Difference Grade, a measure of quality comparable to the Subjective Difference Grade (SDG), which is calculated as the difference between the quality rating of the reference and the test signal. The quality ratings are measured with the five point scale defined in ITU-R BS.1116 and thus the SDG and ODG have a range of [-4;0] where -4 stands for very annoying difference and 0 stands for imperceptible difference between reference and test signal. DI: Distortion Index, a quality measure based on the calculated Model Output Variables(MOVs). It is calculated with a trained neural network. Bandwidth: estimate of the bandwidths of both, the reference and test signal NMR: Noise-To-Mask-Ratio, ratio between the noise (introduced by the codec) and the allowed masking threshold RDF: Relative Disturbed Frames, based on NMR. If the NMR of any frequency band is higher 1.5dB the frame is assumed to be disturbed. The number of disturbed frames is then divided by the number of all frames MFPD: Maximum Filtered Probability of Detection. A detection probability of hearing the noise/artefacts is calculated similar to the NMR. This probability is smoothed over time by a low pass filter of first order. After that the maximum of the resulting value and the MFPD of the preceding frame is chosen. ADB: Average Distorted Block. Similar to RDF, the number of distorted blocks is calculated based on the detection probability (if detection probability > 0.5). Then a measure of how hearable the distortion is is divided by this number. EHS: Harmonic Structure of Error, a measure how tonal the noise signal is. The calculation is based on the autocorrelation of the error spectrum. Modulation Difference: measurement of differences between the envelopes of reference and test signal based on a simple loudness measure and low pass filtering. Noise Loudness: a measure of the loudness of the noise signal. If you understand german, it is recommended to read a paper about ITU-R BS.1398, available at http://www-ft.ee.tu-berlin.de/Publikationen/papers/Tonmeistertagung1998.pdf. 4. Description of Error Messages 1000001 memory allocation failed 2000001 division by zero, please contact [email protected] with the test files 3000001 file open error 3000002 file close error 3000003 file read error 3000004 file write error 3000006 unknown file format 3000007 illegal file format 7000001 missing command line argument 7000002 wrong command line argument 9999999 unknown error 5. The algorithm EAQUAL uses a psychoacoustic model to calculate the signals at the basilar membran, the so called "excitation patterns". The calculation is listening level dependent and uses nonlinear addition of masking patterns. Both input files are processed in the model. The cognitive model is a combination of different algorithms/models to rate the perceptability of noise and distortion added by the coding algorithm. Each of these algorithms yields to a MOV (model output variable); all of them are described above. The average of these MOVs over time and channels are the inputs of a neural network, the output of the network is the DI (distortion index). This value is mapped through a nonlinear (sigmoid) function to the ODG, the final result. 6. License of EAQUAL The source was originally developed at Heinrich-Hertz-Institut fuer Nachrichtentechnik and was then published under the GPL. For more information on the GPL read the license included with the source code package. Please note that the use of this software may require the payment of patent royalties. You need to consider this issue before you start building derivative works. We are not warranting or indemnifying you in any way for patent royalities! YOU ARE SOLELY RESPONSIBLE FOR YOUR OWN ACTIONS! 7. HTML documentation of source code An automatically generated documentation of classes, functions, etc. is available at /EAQUAL/docs/htm/index.htm
About
EAQUAL stands for Evaluation Of Audio Quality. It's an objective measurement technique used to measure the quality of encoded/decoded audio files (very similiar to PEAQ).
Resources
Stars
Watchers
Forks
Releases
No releases published
Packages 0
No packages published
Languages
- C 54.1%
- C++ 25.7%
- Shell 12.5%
- Makefile 5.7%
- HTML 1.9%
- Python 0.1%