There are not many songs that have official recorded versions in multiple languages, but there are lots of websites that offer translations of lyrics and subtitled musical videos on video streaming services. Perhaps one reason for the lack of official versions of popular songs in other languages is “singability.”
Singability is how researchers Haven Kim, Kento Watanabe, Masataka Goto, and Juhan Nam of the Graduate School of Culture Technology in South Korea and The National Institute of Advanced Industrial Science and Technology in Japan refer to the quality of a translated lyric, which affords it “not only the ability of being sung, but also the suitability (easiness) of being sung.”
In a paper published on August 26, 2023, the researchers set out to computationally analyze and evaluate lyric translations using various metrics across three languages: English, Japanese, and Korean. The criteria utilized for their evaluation included syllable counts, phoneme repetition, musical structure, and semantics.
A basic “comprehensive understanding of lyrics that accounts for their musical, linguistic, and cultural elements” also underlined the evaluation framework, which the researchers say is the first one to include Japanese and Korean lyrics.
A Triple-Language Dataset
The researchers created a dataset that included “singable” lyrics obtained from official collections of commercial song lyrics as well as user-translated lyrics from YouTube.
Also included were Google Translate machine-translated (MT) non-singable lyrics where the source was official singable lyrics from 80 English, 162 Japanese, and 161 Korean songs. The dataset included metadata such as the song’s original language. Represented in the dataset are several genres (e.g., Korean and Japanese pop music, animated film music, and musical theater, among others).
For their analysis, the researchers aligned the lyrics of 162 songs line-by-line in the three languages and divided the lyrics into sections, and established that the average number of syllables per sentence in the dataset was 36.18 for English, 50.89 for Japanese, and 40.40 for Korean.
Line Syntactic Imparity, Section Semantic Parity
The researchers based their evaluation on four metrics: line syllable count distance (Dissyl), phoneme repetition similarity (Simpho), musical structure distance (Dismus), and semantic similarity (Simsem). The goal with these metrics was to determine whether the translated lyrics maintained “the integrity of melodies, degree of phoneme repetition, structural coherence, and semantics of the original lyrics.”
Comparing [adapted] singable lyrics to translated non-singable lyrics, the researchers found evidence that semantics are not prioritized in singable lyric translations at the line level, and that semantic connections at the section level are. One of the examples provided in the paper clearly illustrates low syntactic equivalence across the three languages:
In the phrase “I’m going to travel to find the gold,” it is natural in English to mention “I’m going to travel” before “to find the gold.” However, in Japanese and Korean, expressing “to find the gold (金を探しに, 금을 찾으러)” before “I’m going to travel (旅に出る, 떠난다)” is a more typical and natural construction.
Illustrating a case where there may be concepts that are either semantically and/or non-existent in a given language, the paper also includes an example of the song “MIC Drop” by BTS, a Korean pop group. The term “hyodo (효도, taking care of parents)” is omitted in the translated English version posted by a singer on YouTube because there is no direct equivalent for the concept in English. The researchers add the same concept is “effortlessly” conveyed as “koukou (孝行)” in the Japanese version.
Another conclusion drawn was that, to produce an equivalent message across these languages, there would be 41% more syllables in Japanese than in English and 26% more syllables than Korean. This is one reason why translators often modify the meaning of original lyrics.
Lastly, as they had anticipated, the researchers found that phoneme repetition in the original lyrics “is frequently mirrored in the translated lyrics, musically similar sections tend to share the same phonemes and display comparable degrees of phoneme repetition.” They added that some sections, such as choruses, “require a substantial degree of phoneme repetition, while others do not.”