Academia.eduAcademia.edu

Figure 16 – uploaded by Meinard Müller

See full PDFdownloadDownload figure
Table 5.14. Recognition accuracies using features with and without beat synchronization, tested with HMM. Train: DPeatles |) DBeatles | Test: DBeatles | Effect of Using Beat Synchronized Feature

Table 5 14. Recognition accuracies using features with and without beat synchronization, tested with HMM. Train: DPeatles |) DBeatles | Test: DBeatles | Effect of Using Beat Synchronized Feature

Related Figures (61)

Figure 1.1. (a) C major triad played simultaneously, (b) C major triad in arpeggio. A chord is defined as the simultaneous sounding of two or more different notes [21]. Nor- mally the notes which form the chord are played together just like a musician presses the corresponding keys at the same time. In some special cases, these notes are played separately yet successively as a musician press the key one by one. Such cases include arpeggios and broken chords. Figure 1.1 illustrates a chord with the notes played simulta- neously and Figure 1.1(b) illustrates the arpeggio, where the notes are played separately. Chords themselves and their progression are very important because they compose the harmonic content of a music piece. The analysis of harmonic content is essential in the Western tonal music, thus chords as the fundamental component play a crucial role fot the understanding of such music [29]. Furthermore, being a higher-level representation of a music piece compared to the note-level representation, chords bring more insight tc the analysis of the music structure and therefore assist many music information retrieva applications such as cover song identification and music segmentation.
Figure 1.1. (a) C major triad played simultaneously, (b) C major triad in arpeggio. A chord is defined as the simultaneous sounding of two or more different notes [21]. Nor- mally the notes which form the chord are played together just like a musician presses the corresponding keys at the same time. In some special cases, these notes are played separately yet successively as a musician press the key one by one. Such cases include arpeggios and broken chords. Figure 1.1 illustrates a chord with the notes played simulta- neously and Figure 1.1(b) illustrates the arpeggio, where the notes are played separately. Chords themselves and their progression are very important because they compose the harmonic content of a music piece. The analysis of harmonic content is essential in the Western tonal music, thus chords as the fundamental component play a crucial role fot the understanding of such music [29]. Furthermore, being a higher-level representation of a music piece compared to the note-level representation, chords bring more insight tc the analysis of the music structure and therefore assist many music information retrieva applications such as cover song identification and music segmentation.
are “triad” chords. The name “triad” indicates that such chord is composed by three distinct notes. However, since notes which in different octave yet in the same pitch class may be considered as one note when people analyze the chord, it is better to count the number of distinct “pitch classes” or “chromas” instead of distinct notes. Figure 1.3 illustrates how notes are mapped to chroma’. To avoid confusion, in this thesis when we mention a component note of a chord, we actually treat all the notes in a pitch class as one note disregarding their octave information. Figure 1.2. Three types of chords with different number of notes based on root note C. From left to right: C major triad, C dominant seventh and C major ninth.
are “triad” chords. The name “triad” indicates that such chord is composed by three distinct notes. However, since notes which in different octave yet in the same pitch class may be considered as one note when people analyze the chord, it is better to count the number of distinct “pitch classes” or “chromas” instead of distinct notes. Figure 1.3 illustrates how notes are mapped to chroma’. To avoid confusion, in this thesis when we mention a component note of a chord, we actually treat all the notes in a pitch class as one note disregarding their octave information. Figure 1.2. Three types of chords with different number of notes based on root note C. From left to right: C major triad, C dominant seventh and C major ninth.
Figure 1.3. (a) Chromatic circle. (b) Shepards helix of pitch perception [2].
Figure 1.3. (a) Chromatic circle. (b) Shepards helix of pitch perception [2].
As discussed above, the chords are classified by the number of component notes and also by the musical intervals between the notes. Moreover, for each type of chord, we have 12 distinct instances because we have 12 pitch classes which can be used as root note. For example, for the case of a major triad, we have C major triad, C* major triad, D major triad, and so on. Figure 1.6 shows all 12 major triads with the red notes indicating the respective root note. In this thesis, if not specially declared, we will assume that all the chords we consider are triad chords, and we notated them by indicating the root note and type of intervals. For example, C is the short form to denote the C major triad and Cm to denote C minor triad. Figure 1.6. All 12 major triads, the note with red color indicates the respective root note
As discussed above, the chords are classified by the number of component notes and also by the musical intervals between the notes. Moreover, for each type of chord, we have 12 distinct instances because we have 12 pitch classes which can be used as root note. For example, for the case of a major triad, we have C major triad, C* major triad, D major triad, and so on. Figure 1.6 shows all 12 major triads with the red notes indicating the respective root note. In this thesis, if not specially declared, we will assume that all the chords we consider are triad chords, and we notated them by indicating the root note and type of intervals. For example, C is the short form to denote the C major triad and Cm to denote C minor triad. Figure 1.6. All 12 major triads, the note with red color indicates the respective root note
Table 1.1. Ground truth chord labels of Bach BWV846. corresponding ground truth labels are shown in Table 1.1. In Figure 1.8(b), we show the output of a chord recognition system for the classical piece of music Bach BWVS846. The results of four different versions are visualized in our Interpretation Switcher demo, with the top one being the ground truth for MIDI version, and the others being the computed result for three different interpretations played by various performers on different instruments. Each colored interval represents a certain chord and its length corresponds to the duration of that chord. By comparing the difference between any of the three computed results with the ground truth, we can easily find recognition errors such as the 2nd measure of the Koopman’s interpretation, which wrongly recognize Dm represent by pink color as F represent by green color.
Table 1.1. Ground truth chord labels of Bach BWV846. corresponding ground truth labels are shown in Table 1.1. In Figure 1.8(b), we show the output of a chord recognition system for the classical piece of music Bach BWVS846. The results of four different versions are visualized in our Interpretation Switcher demo, with the top one being the ground truth for MIDI version, and the others being the computed result for three different interpretations played by various performers on different instruments. Each colored interval represents a certain chord and its length corresponds to the duration of that chord. By comparing the difference between any of the three computed results with the ground truth, we can easily find recognition errors such as the 2nd measure of the Koopman’s interpretation, which wrongly recognize Dm represent by pink color as F represent by green color.
Figure 1.7. Overview of our chord recognition system.
Figure 1.7. Overview of our chord recognition system.
Figure 1.8. (a) Score of Bach BWV846. (b) corresponding Interpretation Switcher visualization of chord recognition output for the first 11 measures, with the top line being the ground truth, and other three being the computed results.
Figure 1.8. (a) Score of Bach BWV846. (b) corresponding Interpretation Switcher visualization of chord recognition output for the first 11 measures, with the top line being the ground truth, and other three being the computed results.
Figure 1.9. Module of thesis structure.
Figure 1.9. Module of thesis structure.
Figure 2.1. A sample array of filters with their respective magnitude responses in dB. (reproduced from [23]) The decomposition from the original signal to 88 pitch related freq as a prerequisite signal processing step. It provides meaningful uency subbands serve: information about. the signal both in the temporal and the frequency domain, indicating at what time whicl frequency component is present. However, the great contribution o this step only works i a pre-condition is fulfilled: we assumed that the audio file needs to have correct tuning anc the instruments are tuned according to the equal-tempered scale. n this way, tuning is ¢ crucial factor which the multirate filter bank is sensitive to. In this thesis, we previousl; tuned all the audio files before conducting any experiments in order to make sure the pitch features are correct. Besides this, strong onsets or string vibratos will lead to energy spreading in a large frequency range, affecting many subbands and therefore making th decomposition problematic. Without prior detection and smoothing of these frequency fluctuations, one may have imperfect subbands signals when such phenomena occur.
Figure 2.1. A sample array of filters with their respective magnitude responses in dB. (reproduced from [23]) The decomposition from the original signal to 88 pitch related freq as a prerequisite signal processing step. It provides meaningful uency subbands serve: information about. the signal both in the temporal and the frequency domain, indicating at what time whicl frequency component is present. However, the great contribution o this step only works i a pre-condition is fulfilled: we assumed that the audio file needs to have correct tuning anc the instruments are tuned according to the equal-tempered scale. n this way, tuning is ¢ crucial factor which the multirate filter bank is sensitive to. In this thesis, we previousl; tuned all the audio files before conducting any experiments in order to make sure the pitch features are correct. Besides this, strong onsets or string vibratos will lead to energy spreading in a large frequency range, affecting many subbands and therefore making th decomposition problematic. Without prior detection and smoothing of these frequency fluctuations, one may have imperfect subbands signals when such phenomena occur.
Figure 2.2. Score representation, spectrogram and pitch representation of the synthesized music audio ChordExample1.
Figure 2.2. Score representation, spectrogram and pitch representation of the synthesized music audio ChordExample1.
Figure 2.3. Score representation and chromagram of CP extracted from ChordExample1.
Figure 2.3. Score representation and chromagram of CP extracted from ChordExample1.
Figure 2.4. CLP features extracted from ChordExample1. The procedure of how we derived the chroma features with logarithmic compression is presented as follows. Firstly, we compute pitch features from the audio file as we described in Section 2.1. Secondly, the pitch representation is logarithmized by replacing each entry e by the value log(7 -e + 1), where 77 is a suitable positive constant. Thirdly, we convert the logarithmic pitch features into the chroma representation by spectral pooling and @? normalization as we described in Section 2.2. The resulting features are named as Chroma Features with Logarithmic Compression and denoted as CLP(7) where the parameter 7 specifies the extent of logarithmic compression. The role of 7, which is set to 7 = 1000 in most of our experiments, is evaluated in Section 5.3.3.
Figure 2.4. CLP features extracted from ChordExample1. The procedure of how we derived the chroma features with logarithmic compression is presented as follows. Firstly, we compute pitch features from the audio file as we described in Section 2.1. Secondly, the pitch representation is logarithmized by replacing each entry e by the value log(7 -e + 1), where 77 is a suitable positive constant. Thirdly, we convert the logarithmic pitch features into the chroma representation by spectral pooling and @? normalization as we described in Section 2.2. The resulting features are named as Chroma Features with Logarithmic Compression and denoted as CLP(7) where the parameter 7 specifies the extent of logarithmic compression. The role of 7, which is set to 7 = 1000 in most of our experiments, is evaluated in Section 5.3.3.
Figure 2.5. CENS features extracted from ChordExample1.
Figure 2.5. CENS features extracted from ChordExample1.
Figure 2.6. CRP(1,1) extracted from ChordExample1. CRP features have three specialities. Firstly, as we mentioned above they are designed to be invariant with timbre. Secondly, they integrate logarithmic compression in intensity computing. Thirdly, they include the smoothing technique. All these specialities make CRP features outstanding among all feature types. In the following, similar to CENS features we denote CRP features by CRP(w,d). The MATLAB function of how we derived CRP features can be found in pitchSTMSP_to_CRP.m from the Chroma Toolbox|22].
Figure 2.6. CRP(1,1) extracted from ChordExample1. CRP features have three specialities. Firstly, as we mentioned above they are designed to be invariant with timbre. Secondly, they integrate logarithmic compression in intensity computing. Thirdly, they include the smoothing technique. All these specialities make CRP features outstanding among all feature types. In the following, similar to CENS features we denote CRP features by CRP(w,d). The MATLAB function of how we derived CRP features can be found in pitchSTMSP_to_CRP.m from the Chroma Toolbox|22].
Figure 2.7. Tracked frequencies, tracked magnitude, and chroma representation of CISP features extracted from ChordExample1.
Figure 2.7. Tracked frequencies, tracked magnitude, and chroma representation of CISP features extracted from ChordExample1.
Figure 3.1. Binary templates for C and Cm. For example, the binary template corresponding to C C = {C,E, G} is given by
Figure 3.1. Binary templates for C and Cm. For example, the binary template corresponding to C C = {C,E, G} is given by
Figure 3.2. Harmonically enriched templates for C and Cm. Table 3.2. Contributions for the first six harmonics of component notes of C. Assuming the chords is played in the 4th octave. From left to right corresponding to the notes C4, E4 and G4. templates we describe in this set are not fixed as the previous binary or harmonically enriched setting, but vary according to different training datasets. These are the most significant differences compared to the previous two settings.
Figure 3.2. Harmonically enriched templates for C and Cm. Table 3.2. Contributions for the first six harmonics of component notes of C. Assuming the chords is played in the 4th octave. From left to right corresponding to the notes C4, E4 and G4. templates we describe in this set are not fixed as the previous binary or harmonically enriched setting, but vary according to different training datasets. These are the most significant differences compared to the previous two settings.
Figure 3.3. Procedure of generating the averaged template 7”. Note that we shift major chords to C and minor chords to Cm .
Figure 3.3. Procedure of generating the averaged template 7”. Note that we shift major chords to C and minor chords to Cm .
Table 3.3. Mapping function d¢po-q for major and minor triads.
Table 3.3. Mapping function d¢po-q for major and minor triads.
Figure 3.4. Averaged templates of different features for chord C and Cm.
Figure 3.4. Averaged templates of different features for chord C and Cm.
Figure 3.5. Three variation of template sets. (a)T°. (b)7”. (c)T° using feature CENS(1, 1). It is hard to judge the quality of these methods by whether they are set with theoretically values or practical values. Sometimes the recognition method using simple binary tem- plates performs the best, and the harmonically enriched templates or averaged templates can also win in some test experiments. However there is a problem with the averaged tem- plates. They are heavily depending on the training data. If the training data contains too many bad instances of chords, the templates we computed might be biased. For example if a training dataset contains more C augmented triad than C, since in our system the augmented are mapped to major, the resulting template will have a higher weight both to the note G and G?, which in theory G# should not occur, and of course such kind of template is not the exact pattern of C.
Figure 3.5. Three variation of template sets. (a)T°. (b)7”. (c)T° using feature CENS(1, 1). It is hard to judge the quality of these methods by whether they are set with theoretically values or practical values. Sometimes the recognition method using simple binary tem- plates performs the best, and the harmonically enriched templates or averaged templates can also win in some test experiments. However there is a problem with the averaged tem- plates. They are heavily depending on the training data. If the training data contains too many bad instances of chords, the templates we computed might be biased. For example if a training dataset contains more C augmented triad than C, since in our system the augmented are mapped to major, the resulting template will have a higher weight both to the note G and G?, which in theory G# should not occur, and of course such kind of template is not the exact pattern of C.
together in a dataset and name it as D*. See Table 5.1 for details. The chord annotations in this dataset are precisely written by Verena Konz. Furthermore, we particularly choose 16 Beatles recordings and four classical pieces together as a combination of pop and classical music. We name it as D?°. See Table 5.2 for detail. Table 5.1. Description of the dataset D*.
together in a dataset and name it as D*. See Table 5.1 for details. The chord annotations in this dataset are precisely written by Verena Konz. Furthermore, we particularly choose 16 Beatles recordings and four classical pieces together as a combination of pop and classical music. We name it as D?°. See Table 5.2 for detail. Table 5.1. Description of the dataset D*.
Table 5.2. Description of the dataset D?°.
Table 5.2. Description of the dataset D?°.
From Table 5.3 we can already tell the huge recognition difference caused by feature side. For all the chord recognizers, the best feature is always the CRP(13,1) and the worst is CP for most of the time and CISP for one time. And for each of the recognizers, the biggest increased recognition accuracy is 0.20, 0.31, 0.25, 0.34, 0.26 and 0.21 respectively. This implies the importance of computing audio features in a suitable way. Without good features, one can not obtain satisfiable result even using the complex chord recognizers.
From Table 5.3 we can already tell the huge recognition difference caused by feature side. For all the chord recognizers, the best feature is always the CRP(13,1) and the worst is CP for most of the time and CISP for one time. And for each of the recognizers, the biggest increased recognition accuracy is 0.20, 0.31, 0.25, 0.34, 0.26 and 0.21 respectively. This implies the importance of computing audio features in a suitable way. Without good features, one can not obtain satisfiable result even using the complex chord recognizers.
Figure 5.1. Bar visualization of Table 5.3. Result comparison from feature side. e . Recognition accuracles of all f atures and all T cognition method: . D . Test: D © e . ° ° e Ss. Train © 1
Figure 5.1. Bar visualization of Table 5.3. Result comparison from feature side. e . Recognition accuracles of all f atures and all T cognition method: . D . Test: D © e . ° ° e Ss. Train © 1
Ta le 5. 4, COg_gnit1lo Re : ° ¢. nN ac curacies of all feat ures wit a ‘0. oO et Ss. Nn: rec eniti h od al b iti h I De a
Ta le 5. 4, COg_gnit1lo Re : ° ¢. nN ac curacies of all feat ures wit a ‘0. oO et Ss. Nn: rec eniti h od al b iti h I De a
Table 5.5 shows the evaluation result using four chord recognition methods and features with and without logarithmic compression. Here, CP denotes the original pitch features without logarithmic compression. Other features are CLP with different extent of compres- sion specified by 7. Besides, Figure 5.3 is generated by using the numbers of Table 5.5 and offers a better visualization.
Table 5.5 shows the evaluation result using four chord recognition methods and features with and without logarithmic compression. Here, CP denotes the original pitch features without logarithmic compression. Other features are CLP with different extent of compres- sion specified by 7. Besides, Figure 5.3 is generated by using the numbers of Table 5.5 and offers a better visualization.
Figure 5.3. Visualization of Table 5.5. Effect of logarithmic compression. Table 5.5. Recognition differences of features with various logarithmic compression parameter. Train: DPeatles, Tost; D Beatles,
Figure 5.3. Visualization of Table 5.5. Effect of logarithmic compression. Table 5.5. Recognition differences of features with various logarithmic compression parameter. Train: DPeatles, Tost; D Beatles,
Figure 5.4. Comparisons of different logarithmic compression on chroma features
Figure 5.4. Comparisons of different logarithmic compression on chroma features
Figure 5.5. Recognition accuracies of using different smoothing window length in CENS features. Train: pPowles, Test: Ty pesvies
Figure 5.5. Recognition accuracies of using different smoothing window length in CENS features. Train: pPowles, Test: Ty pesvies
Figure 5.6. Recognition accuracies of using different smoothing window length in CRP features. Train: DPols, Tesh, DEES,
Figure 5.6. Recognition accuracies of using different smoothing window length in CRP features. Train: DPols, Tesh, DEES,
(a) Score of the first 10 seconds (corresponding to the first three measures) of BeatlesLetItBe.
(a) Score of the first 10 seconds (corresponding to the first three measures) of BeatlesLetItBe.
Figure 5.7. Comparisons of different window lengths of smoothing on CENS features. Test « recognizer T> and test data : the first 10 seconds of BeatlesLetItBe.
Figure 5.7. Comparisons of different window lengths of smoothing on CENS features. Test « recognizer T> and test data : the first 10 seconds of BeatlesLetItBe.
Table 5.6. Recognition differences of pitch range separation of CISP. Train: DP<@tles, Test: DBeatles between them is also larger than those tested on CISP features. A common phenomenon occurs in both CLP and sCISP is the bad performance of using bass range features. Again we get about 7% to 20% decrease.
Table 5.6. Recognition differences of pitch range separation of CISP. Train: DP<@tles, Test: DBeatles between them is also larger than those tested on CISP features. A common phenomenon occurs in both CLP and sCISP is the bad performance of using bass range features. Again we get about 7% to 20% decrease.
For the dimension comparison of CLP, the result varies with different compression param- eters. For CLP(1), the “24dim all” features improves recognition at most 5% and at least 2%. However, this improvement becomes less and less as the log compression parameter 17 gets larger. Until 7 = 100 it still improves the result. But to our surprise, when 7 = 100, it decreases the result of GP by 30% and HMM by 35%, while it stays nearly the same for T° and T?. This is because for large compression parameters, the intensity difference between chromas is already quite small and hard to distinguish, now putting such undistinguished chroma features into a complex model, the model pattern we learned will be even more undistinguished among each of the chords. That is why for the template method the performance is still ok but for complex methods it totally fails.
For the dimension comparison of CLP, the result varies with different compression param- eters. For CLP(1), the “24dim all” features improves recognition at most 5% and at least 2%. However, this improvement becomes less and less as the log compression parameter 17 gets larger. Until 7 = 100 it still improves the result. But to our surprise, when 7 = 100, it decreases the result of GP by 30% and HMM by 35%, while it stays nearly the same for T° and T?. This is because for large compression parameters, the intensity difference between chromas is already quite small and hard to distinguish, now putting such undistinguished chroma features into a complex model, the model pattern we learned will be even more undistinguished among each of the chords. That is why for the template method the performance is still ok but for complex methods it totally fails.
Table 5.8. Recognition differences between various recognizers, using features without smoothing. Trains DPestes, Tegy, DPeattes, .4.1 Overall Performance
Table 5.8. Recognition differences between various recognizers, using features without smoothing. Trains DPestes, Tegy, DPeattes, .4.1 Overall Performance
Figure 5.8. Visualization of Table 5.8. Overall performance of the recognizers.
Figure 5.8. Visualization of Table 5.8. Overall performance of the recognizers.
Figure 5.9. Error analysis of chord recognizers. Upper: major chord. Lower: minor chords. Left: on template based methods. Right: on statistical model based methods.
Figure 5.9. Error analysis of chord recognizers. Upper: major chord. Lower: minor chords. Left: on template based methods. Right: on statistical model based methods.
Figure 5.10. Comparisons of having different suppression parameter on off-diagonal values in trained covariance matrix.
Figure 5.10. Comparisons of having different suppression parameter on off-diagonal values in trained covariance matrix.
Figure 5.11. Comparisons of having different scale on diagonal values in trained covariance matrix.
Figure 5.11. Comparisons of having different scale on diagonal values in trained covariance matrix.
Figure 5.12. Comparisons of having different scale on all the diagonal values and off-diagonal ai component notes in trained covariance matrix.
Figure 5.12. Comparisons of having different scale on all the diagonal values and off-diagonal ai component notes in trained covariance matrix.
Figure 5.13. Comparisons of having different scale on all the diagonal values and off-diagonal at component notes in manually set covariance matrix.
Figure 5.13. Comparisons of having different scale on all the diagonal values and off-diagonal at component notes in manually set covariance matrix.
Figure 5.14. Trained covariance matrix of corresponding feature types and manually set covari- ance matrix for C.
Figure 5.14. Trained covariance matrix of corresponding feature types and manually set covari- ance matrix for C.
Figure 5.15. Transition Matrix. However, in HMM we are not selecting the maximum, and the value of Gaussian probabilities severely affects the final performance. This is because the Viterbi algorithm takes bott the covariance matrix and transition matrix and then computes the forward probability in an accumulated multiplication manner. In this way, any tiny difference even as smal as 0.001 will be exaggerated largely after several times of multiplication. Therefore, if one wants to use the manually set covariance matrix in HMM, one needs to be very carefully tc scale the values to fit the intensity of the features. Figure 5.14(a), 5.14(b) and 5.14(c) show the example of covariance feature setting of three features. We can see that the precision of the color bars of CISP and CLP(10) is at 10~% but CENS is at 10~?. This is the reason why the trends of CISP and CLP(10) are close to each other but different witl CENS in Figure 5.13(b). Furthermore, by this experiment, we see that using the manually set covariance matrix can also yield a good recognition result as long as the scale of the values are appropriately set.
Figure 5.15. Transition Matrix. However, in HMM we are not selecting the maximum, and the value of Gaussian probabilities severely affects the final performance. This is because the Viterbi algorithm takes bott the covariance matrix and transition matrix and then computes the forward probability in an accumulated multiplication manner. In this way, any tiny difference even as smal as 0.001 will be exaggerated largely after several times of multiplication. Therefore, if one wants to use the manually set covariance matrix in HMM, one needs to be very carefully tc scale the values to fit the intensity of the features. Figure 5.14(a), 5.14(b) and 5.14(c) show the example of covariance feature setting of three features. We can see that the precision of the color bars of CISP and CLP(10) is at 10~% but CENS is at 10~?. This is the reason why the trends of CISP and CLP(10) are close to each other but different witl CENS in Figure 5.13(b). Furthermore, by this experiment, we see that using the manually set covariance matrix can also yield a good recognition result as long as the scale of the values are appropriately set.
Figure 5.16. Comparisons of different suppression parameters on diagonal values of a transition matrix. In this subsection, we inspect the diagonal entries of transition matrix. Figure 5.15 shows the transition matrix trained on DBe*"!*s, All the diagonal values are above 0.95. They are much larger than other off-diagonal values, and most of them are below 0.001. A diagonal entry represents the self transition of a chord, and an off-diagonal entry represents the
Figure 5.16. Comparisons of different suppression parameters on diagonal values of a transition matrix. In this subsection, we inspect the diagonal entries of transition matrix. Figure 5.15 shows the transition matrix trained on DBe*"!*s, All the diagonal values are above 0.95. They are much larger than other off-diagonal values, and most of them are below 0.001. A diagonal entry represents the self transition of a chord, and an off-diagonal entry represents the
5.5 Effect of Tuning
5.5 Effect of Tuning
Figure 5.17. Histogram of recognition with tuning and without tuning, using recognizer T°. Table 5.9. Recognition differences between features with tuning and without tuning. Train: DPertles | Test: DBeatles Although it is asserted by people that the audio files which prepared for chord recognition are in standard tuning, which means for example in a concert the note A4 should be exactly 440 Hz, the actual tuning of the music recordings sometimes deviate from the standard tuning. In this section, we check the influence of tuning on the chord recognition. To this
Figure 5.17. Histogram of recognition with tuning and without tuning, using recognizer T°. Table 5.9. Recognition differences between features with tuning and without tuning. Train: DPertles | Test: DBeatles Although it is asserted by people that the audio files which prepared for chord recognition are in standard tuning, which means for example in a concert the note A4 should be exactly 440 Hz, the actual tuning of the music recordings sometimes deviate from the standard tuning. In this section, we check the influence of tuning on the chord recognition. To this
Table 5.11. Recognition differences between features with and without HPSS. Train: DBe@tles. Test: DBeatles |
Table 5.11. Recognition differences between features with and without HPSS. Train: DBe@tles. Test: DBeatles |
Table 5.10. Recognition accuracies using different training datasets. Result test on DPtles, (a)using feature CP. (b)CENS(1,1). (c)CRP(1, 1).
Table 5.10. Recognition accuracies using different training datasets. Result test on DPtles, (a)using feature CP. (b)CENS(1,1). (c)CRP(1, 1).
Table 5.12. Maximum increased recognition difference with and without HPSS.
Table 5.12. Maximum increased recognition difference with and without HPSS.
Figure 5.18. Histogram of recognition difference with HPSS, using feature CLP(1000). Table 5.13. Maximum decreased recognition difference with and without HPSS.
Figure 5.18. Histogram of recognition difference with HPSS, using feature CLP(1000). Table 5.13. Maximum decreased recognition difference with and without HPSS.
Table 5.15. Recognition accuracies using features with and without beat synchronization , tested with qT. Train: Beatles U py Deaues Test: DBeatles |
Table 5.15. Recognition accuracies using features with and without beat synchronization , tested with qT. Train: Beatles U py Deaues Test: DBeatles |
Table 5.16. Recognition accuracies of classical dataset using T®. Test: D4 ) Experiments on Classical Dataset
Table 5.16. Recognition accuracies of classical dataset using T®. Test: D4 ) Experiments on Classical Dataset
Table 5.17. Recognition accuracies of classical dataset using HMM. Train: DP«*les, Test: D4
Table 5.17. Recognition accuracies of classical dataset using HMM. Train: DP«*les, Test: D4
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved