Jingju a Cappella Recordings Collection

The Jingju a Cappella Recordings Collection (JaCRC) is part of the Jingju Music Corpus created in the CompMusic project at the Music Technology Group, Universitat Pompeu Fabra, Barcelona (MTG). The JaCRC was created for different research tasks, mostly concerning melodic characteristics of jingju arias and pronunciation in jingju, and parts of the collection have been used in several publications. The JaCRC contains 314 recordings of jingju a cappella singing, plus 76 recordings of the jinghu accompaniment for their corresponding vocal tracks. Except for 53 of them (see CONTENT below), all of the recordings were newly created for this collection. The JaCRC also contains the manual segmentation of 217 vocal recordings and lyrics files for 156, 67 of which include annotations for start and end of each lyrics line in a related music score (see the README file). The dataset is released under a Creative Commons license (see LICENSE below).

The content of the JaCRC was previously published in three different parts (part 1, part 2, part 3). This new release puts all the data together under an unified structure in order to ease its usability.

CONTENT

The main body of the JaCRC are 239 a cappella recordings of jingju arias. Among those, the main contribution of the collection are the 186 newly created a cappella recordings by professional or semi-professional actors. Some of the recordings contain incomplete arias because the performer decided to stop according to their own will. The aria is then completed in subsequent recording(s). In few occasions, the performer decided to record a second version of the same aria. Both versions are included in the collection.

The performers for 76 of these recordings sung over a jinghu accompaniment played live in a different room. These accompaniments were also recorded and added to the JaCRC.

To complement the collection, recordings from existing sources were also integrated to the JaCRC. 15 a cappella recordings were obtained from commercial releases by subtracting the instrumental accompaniment, published in separate tracks to be used as accompaniment by amateur singers, from the mixed track. These recordings are not included in the JaCRC for copyright issues, but can be shared for research purposes only (see CONTACT below). However, the metadata and the segmentation files for these 15 recordings have been included in the JaCRC. Besides, 53 a cappella jingju recordings from Singing Voice Audio Dataset were included here with permission of their authors (see LICENSE and USE below).

With the goal of developing technologies to aid learning of jingju singing, 75 recordings were created from amateur performers, both children and adults. These amateur performers, considered as ‘students,’ sung trying to imitate a reference model, considered as ‘teacher.’ The ‘teacher’ would be either present in the session, and their performances were also recorded, or an existing recording of the JaCRC was played as model. The 16 recordings of the teachers are part of the JaCRC and the anonymized recordings of the students are included in the JaCRC.

All the artists recorded for the JaCRC manifested their written consent to the MTG for the public release of these recordings under Creative Common license.

In order to be used for different research tasks, 142 recordings were manually segmented to the phrase and syllable level. Among these, 81 recordings, including those 16 ones used as ‘teacher’ recordings, were further segmented to the phoneme level. All ‘student’ recordings were also segmented to the phrase, syllable and phoneme level. These segmentations are included in the JaCRC as Praat TextGrid files.

For 156 recordings there are corresponding csv files containing the lyrics performed in the recording, one line per row. Among these, 67 csv files also contain annotations for the boundaries of each lyrics line in a related music score. The boundaries are annotated as offset according to the music21 toolkit. The related music scores can be found in the Jingju Music Scores Collection with the same name as the one annotated in the csv files.

COVERAGE

As part of the Jingju Music Corpus, the JaCRC was gathered with the purpose of studying the most representative characteristics of jingju vocal music, and therefore the most representative instances of the main elements of jingju vocal music, that is, role type, shengqiang and banshi, are well covered in the collection. Below some statistics about the coverage of these elements in the JaCRC are given. The numbers in brackets correspond to the number of recordings that include (not always exclusively) that element and its percentage with respect to the total 254 recordings in the collection. The numbers include the 15 recordings from commercial realeases not available in the collection (see CONTENT above).

Regarding role types, the JaCRC includes 5 different ones. The two most extensively covered ones are dan (127, 50.0%), including male dan (27) and huadan (2), and laosheng (108, 42.5%), including female laosheng (8). The other role types included in the JaCRC are jing (17, 6.7%), most of them of female jing (16), xiaosheng (1, 0.4%) and chou (1, 0.4%).

The two main shengqiang in jingju are extensively covered in the JaCRC, namely xipi (153, 60.2%) and erhuang (62, 24.4%). Besides, other 7 shengqiang are also present in the collection, namely sipingdiao (14, 5.5%), nanbangzi (11, 4.3%), fan’erhuang (8, 3.1%), fansipingdiao (2, 0.8%), fanxipi (4, 1.6%), gaobozi (1, 0.4%), and handiao (1, 0.4%).

As for banshi, there are instances of 18 different ones included in the JaCRC. The 7 more extensively represented banshi are yuanban (76, 29.9%), liushui (63, 24.8%), manban (46, 18.1%), erliu (40, 15.7%), sanban (34, 13.4%), yaoban (34, 13.4%), and daoban (27, 10.6%). Other banshi also included in the collection are kuaiban (17, 6.7%), huilong (8, 3.1%), sanyan (7, 2.8%), kuaisanyan (7, 2.8%), mansanyan (3, 1.2%), zhongsanyan (3, 1.2%), pengban (2, 0.8%), gunban (1, 0.4%), duoban (1, 0.4%), shuban (1, 0.4%), and kuaisanban (1, 0.4%).

In terms of content, the JaCRC contains recordings of 142 arias from 74 different plays.

Finally, the recordings in the JaCRC are performed by 23 artists, including 8 professional actors, 2 graduated jingju students, 3 undergraduate jingju students in their 4th year, and 10 amateur performers. In terms of role types, there are 10 laosheng performers, one of them being the one who also performs the xiaosheng and jing recordings, and another one also performing the chou recording, 8 dan, one of them also performing the huadan recordings, 3 male dan, 1 female laosheng and 1 female jing.

ANNOTATIONS

All the annotation files are named in the same exact manner as its corresponding recording, so that they can be easily matched. Besides, the metadata and information csv files indicate which annotations are available for which recordings.

There are two types of annotations: segmentation and lyrics.

The segmentation annotations were done manually and in three phases, corresponding to the subfolders in the “JaCRC-annotations” folder numbered ‘1,’ ‘2’ and ‘3.’ All the segmentations were done using the software Praat and are available in the JaCRC as TextGrid files. The phoneme annotations follow the Extended Speech Assessment Methods Phonetic Alphabet (X-SAMPA). Below is a description of the annotations contained in each of the subfolders:

“1-phrase-syllable-phoneme” folder: all the recordings whose annotations are contained in this folder were segmented at least to the phrase (lyrics line), syllable and phoneme levels. Since the annotations were done for different research tasks, the TextGrid files might contain different numbers of tiers, but all of them have a tier named ‘line’ for the phrase level segmentation with lyrics line in Chinese characters as labels, a tier named ‘pinyin’ for the syllable level segmentation with syllables in the pinyin romanization system as labels, and a ‘details’ tier for phoneme segmentation and labels in X-SAMPA. In order to ease access to these annotations, tab-separated values files were generated from the TextGrid files and also included as txt files in this folder. The files that add “_phrase” to the recording’s name contain the phrase level annotations in pinyin. Those that add “_phrase_char” contain the same phrase level annotations, but in Chinese characters. Those that add “_syllable” contain the syllable level annotations in pinyin. And those that add “_phoneme” contain the phoneme level annotations in X-SAMPA.

“2-phrase-syllable” folder: same case as in the previous folder, but without phoneme level annotations. In these TextGrid files, the phrase level annotations are still in tiers named ‘line,’ and the syllable level ones are in tiers named ‘dianSilence.’

“3-students” folder: same case as in “1-phrase-syllable-phoneme” folder. In these TextGrid files, the phrase level annotations are still in tiers named ‘line,’ the syllable level ones are in tiers named ‘dianSilence,’ and the phoneme level ones in tiers named ‘details.’

The lyrics annotations consist of csv files (semicolon as separator) containing the lyrics of their corresponding recordings in their original Chinese script. Each row corresponds to a lyrics line. The first three columns contain information for “Role type,” “Shengqiang” and “Banshi” (see the README file). In the fourth one, under the heading “Couplet line,” “s” (from shangju) indicates that the corresponding lyrics line is an opening line, “x” (from xiaju) indicates that it is a closing line, and “k” indicates is a kutou line. The fifth column, “Lyrics line,” contains the lyrics. If there is a matching music score in the Jingju Music Scores Collection (JMSC) for the aria performed in the corresponding recording, the sixth column, “Matched score lyrics line,” contains the lyrics for the same as they appear in the score. The seventh column, “Score XML” contains the name of the music score file in the JMSC. Finally, the eight and ninth columns, “Start” and “End,” contain the starting and ending boundaries of the lyrics line in the score. The boundaries are given as note offsets, according to music21. With this information, the notation of each line can be retrieved from the score.

For a thorough description of the JaCRC, including metadata and information, naming convention and sources, please see the README file.

LICENSE

All the recordings newly created for the JaCRC, that is, all of them except for those from the Singing Voice Audio Dataset, are published under a Creative Commons Attribution 4.0 International License.

For the license of the recordings from the Singing Voice Audio Dataset (those whose source in the metadata and information csv files is “SVAD”), included in the JaCRC with permission of the authors, please refer to its website.

REFERENCING THE JaCRC

If you use the recordings of the JaCRC in your research, please reference it in your publications using the text proposed in this website in the section “Cite as.”

If you use the recordings from the Singing Voice Audio Dataset (those whose source in the metadata and information csv files is “SVAD”), please also include the following reference in your publications:

Dawn A. A. Black, Ma Li and Mi Tian. "Automatic Identification of Emotional Cues in Chinese Opera Singing", in Proc. of 13th Int. Conf. on Music Perception and Cognition and the 5th Conference for the Asian-Pacific Society for Cognitive Sciences of Music (ICMPC 13-APSC0M 5 2014), Seoul, South Korea, August 2014.

CONTACT

For more information, or to request access to the recordings from commercial sources, that can be shared only for research purposes, please contact Rafael Caro Repetto (rafael.caro at upf.edu).

ACKNOWLEDGEMENTS

We express our deepest gratitude to all the professional and amateur performers who so generously contributed with their time and their art to the JaCRC.

The creation of the JaCRC was funded by the European Research Council under the European Union’s Seventh Framework Program (FP7/2007-2013), as part of the CompMusic project (ERC grant agreement 267583).