Academia.eduAcademia.edu

Figure 2 – uploaded by Katunobu Itou

See full PDFdownloadDownload figure
In the commentary, to assist interpretation, graphical nota- tion of the melodic line is used. However, this notation does not assist understanding of the melodic line in the way it does in Western musical scores, because of the difficulty in estimating the exact line from acoustic signals. The commentary includes the graphical notation of 300 phrases from 55 pieces. Figure | shows an example of a graphical notation [7] and Figure 2 shows a pitch (f0) contour of its execution. This sample is a phrase in the dynamic mode.

Figure 2 In the commentary, to assist interpretation, graphical nota- tion of the melodic line is used. However, this notation does not assist understanding of the melodic line in the way it does in Western musical scores, because of the difficulty in estimating the exact line from acoustic signals. The commentary includes the graphical notation of 300 phrases from 55 pieces. Figure | shows an example of a graphical notation [7] and Figure 2 shows a pitch (f0) contour of its execution. This sample is a phrase in the dynamic mode.

Related Figures (8)

Figure 2. Pitch contours of Noh singing In Figure 1, the first row is the graphical notation. Bullets show onsets, and lines and curves show pitch transience. In this notation, the different vertical positions indicate different pitches similar to Western music notation. However, this notation is continuous unlike the discrete Western notation. Therefore, ornaments indicated using smaller symbols, e.g, a grace note, are indicated using the combination of curves. The second row contains the lyrics in Japanese kana. The last row contains note names. Figure 2 illustrates the difficulty of finding onsets and transitions of notes because of the extent of certain vibrato being greater than the transition pitch difference. By contrast, the extent of vibratos in Western music hardly exceeds 200 cents. Cent is a unit for musical pitch. One semitone is 100 cents. In this paper, n cents for an fHz pitch is n = 1200 - logs iE: As seen in this example,fitting a melodic line to a pitch contour is difficult because a new melodic line notation, closer to the acoustic signal, is required. Figure 1. A graphical notation of a melodic information in a Noh vocal book
Figure 2. Pitch contours of Noh singing In Figure 1, the first row is the graphical notation. Bullets show onsets, and lines and curves show pitch transience. In this notation, the different vertical positions indicate different pitches similar to Western music notation. However, this notation is continuous unlike the discrete Western notation. Therefore, ornaments indicated using smaller symbols, e.g, a grace note, are indicated using the combination of curves. The second row contains the lyrics in Japanese kana. The last row contains note names. Figure 2 illustrates the difficulty of finding onsets and transitions of notes because of the extent of certain vibrato being greater than the transition pitch difference. By contrast, the extent of vibratos in Western music hardly exceeds 200 cents. Cent is a unit for musical pitch. One semitone is 100 cents. In this paper, n cents for an fHz pitch is n = 1200 - logs iE: As seen in this example,fitting a melodic line to a pitch contour is difficult because a new melodic line notation, closer to the acoustic signal, is required. Figure 1. A graphical notation of a melodic information in a Noh vocal book
TABLE I. SCORE INFORMATION In Western music, pitch transition occurs at the onset. In Noh singing, pitch transition does not occur quickly at the onset, and the execution is like a grace note [1]. In addition, unlike Western music or Japanese popular music, consonants last longer. Reflecting these aspects, a note must be divided into phones, i.e., a note must contain a phone boundary. Each note corresponds to a single syllable. In Japanese, syllable is classified in two types: CV, which means a conso- nant succeeding a vowel, V, which means just a vowel. Hence, each note has one or two phone fields. In addition, each phone field has pitch transition information, which consists of three values at most. The first is an original and mandatory pitch. The second is a first-transited pitch, and the third is a second- transited pitch. Pitch is expressed as an integer whose value difference refers approximately to a halftone of Western music. Table I is an example of the first three notes in Figure 1.
TABLE I. SCORE INFORMATION In Western music, pitch transition occurs at the onset. In Noh singing, pitch transition does not occur quickly at the onset, and the execution is like a grace note [1]. In addition, unlike Western music or Japanese popular music, consonants last longer. Reflecting these aspects, a note must be divided into phones, i.e., a note must contain a phone boundary. Each note corresponds to a single syllable. In Japanese, syllable is classified in two types: CV, which means a conso- nant succeeding a vowel, V, which means just a vowel. Hence, each note has one or two phone fields. In addition, each phone field has pitch transition information, which consists of three values at most. The first is an original and mandatory pitch. The second is a first-transited pitch, and the third is a second- transited pitch. Pitch is expressed as an integer whose value difference refers approximately to a halftone of Western music. Table I is an example of the first three notes in Figure 1.
Figure 3. Phone segmentation of pitch contours Figure 3 exemplifies the phonetic segmentation result of the pitch contour of the first 4 s of Figure 2. Vertical lines indicate onsets. This figure was plotted using manual labeling for a later explanation.
Figure 3. Phone segmentation of pitch contours Figure 3 exemplifies the phonetic segmentation result of the pitch contour of the first 4 s of Figure 2. Vertical lines indicate onsets. This figure was plotted using manual labeling for a later explanation.
Figure 4. An initial centered pitch contour (dot plot) and an initial centered curve (solid line) The pitch for all its segments is not staying, and thus, its mean cannot be calculated. For such pitches, the mean is calculated using the value of the nearest calculated pitch by subtracting the default scale difference value. The scale value is correspondent with 100 cents. For example, in Table I, if 49 was not calculated, ji19 is calculated as p19 = pug * 209-9/12,
Figure 4. An initial centered pitch contour (dot plot) and an initial centered curve (solid line) The pitch for all its segments is not staying, and thus, its mean cannot be calculated. For such pitches, the mean is calculated using the value of the nearest calculated pitch by subtracting the default scale difference value. The scale value is correspondent with 100 cents. For example, in Table I, if 49 was not calculated, ji19 is calculated as p19 = pug * 209-9/12,
Figure 6. Comparison of the transcriptions of different singers of peak similar to float after the dips at approximately 6 s in both transcriptions. Asymmetric vibrato is one of the most specific characteristics of Noh singing, and is considered as an unprescribed ornament. This suggests that new ornaments will be discovered using this method.
Figure 6. Comparison of the transcriptions of different singers of peak similar to float after the dips at approximately 6 s in both transcriptions. Asymmetric vibrato is one of the most specific characteristics of Noh singing, and is considered as an unprescribed ornament. This suggests that new ornaments will be discovered using this method.
Figure 5. A melody transcription. The blue line is f0 contour. The red line is the melody transcribed with the onset timing using the bullets.
Figure 5. A melody transcription. The blue line is f0 contour. The red line is the melody transcribed with the onset timing using the bullets.
Figure 7. A melody transcription using erroneous segmentation Regarding the phone segmentation accuracy, the absolute error compared with the hand-labeled onset was 0.37 s by the general HMM and 0.089 s by the adapted HMMs. Fig- ure 7 shows an example of melody transcription using highly erroneous phone segmentation from the upper pitch contour in Figure 6. The average absolute onset error was 0.23 s. Comparing the data in Figure 6, the first float was missing and the transient just after 6 s was missing. For both cases, successive vowels caused segmentation errors, for example the error of the vowel in the first float was 1.56 s and the error near the transient was 0.39 s, due to the phonation difference in modern Japanese speech. Such vowel differences were reduced by MLLR adaptation. We evaluated this type of transcription error using the average absolute pitch error and compared it to the transcrip- tion estimated using manually-labeled phone segmentation in the cent domain. The average absolute pitch error of the transcription in Figure 7 was 50 cents. The average absolute pitch error was 24 cents by the general HMM segmentation and 13 cents by the adapted HMM segmentation. Figure 8 shows the relationship between the phone segmentation error and the transcription estimation error. The adapted model did not estimate erroneous transcription. For modern or Western music, the onset detection accuracy is evaluated within 50 ms window [10], but for Noh singing more broader window, e.g., 200 ms, might be appropriate for evaluation, because Noh singing is very slow, e.g., the average phone duration was approximately 400 ms in the evaluation data, and the onset is flexible and not as important as it is in Western music.
Figure 7. A melody transcription using erroneous segmentation Regarding the phone segmentation accuracy, the absolute error compared with the hand-labeled onset was 0.37 s by the general HMM and 0.089 s by the adapted HMMs. Fig- ure 7 shows an example of melody transcription using highly erroneous phone segmentation from the upper pitch contour in Figure 6. The average absolute onset error was 0.23 s. Comparing the data in Figure 6, the first float was missing and the transient just after 6 s was missing. For both cases, successive vowels caused segmentation errors, for example the error of the vowel in the first float was 1.56 s and the error near the transient was 0.39 s, due to the phonation difference in modern Japanese speech. Such vowel differences were reduced by MLLR adaptation. We evaluated this type of transcription error using the average absolute pitch error and compared it to the transcrip- tion estimated using manually-labeled phone segmentation in the cent domain. The average absolute pitch error of the transcription in Figure 7 was 50 cents. The average absolute pitch error was 24 cents by the general HMM segmentation and 13 cents by the adapted HMM segmentation. Figure 8 shows the relationship between the phone segmentation error and the transcription estimation error. The adapted model did not estimate erroneous transcription. For modern or Western music, the onset detection accuracy is evaluated within 50 ms window [10], but for Noh singing more broader window, e.g., 200 ms, might be appropriate for evaluation, because Noh singing is very slow, e.g., the average phone duration was approximately 400 ms in the evaluation data, and the onset is flexible and not as important as it is in Western music.
Figure 8. The relationship between the segmentation error and the transcription error
Figure 8. The relationship between the segmentation error and the transcription error
Related topics:
MusicArt
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved