As I belong to a choir that sings Gregorian chants for Mass and different liturgical providers at St. Gregory Home, Institute for Non secular Music in Tokyo, I used to be tremendous curious to be taught extra concerning the mission, so I contacted the Regensburger Domspatzen.
The scholars who developed the AI have been members of the choir since they had been 10 years outdated (fifth grade) and have grown up immersed in numerous types of sacred Christian music, together with Gregorian chants. Though Gregorian chants have a protracted historical past and a ton of cultural heritage, nowadays they’re one thing actually solely deeply recognized by a restricted variety of specialists. Due to films, many individuals will know one thing is a Gregorian chant in the event that they hear one, however few will know the specifics like which one it’s or the way it differs from one other.
Group photograph of the Regensburger Domspatzen. © Regensburger Domspatzen
I requested the scholars how they bought into this and Andreas Haberl, the coed chargeable for growing the code, informed me that the mission started with the query of how machine studying strategies can contribute to preserving the custom of Gregorian chants, “not as an alternative choice to human apply, however as a complementary instrument.”
Andreas informed me that instructing a pc to provide an authentic Gregorian chant is an actual problem. Gregorian chants originated earlier than the ninth century CE and are considered significantly tough to mannequin utilizing machine studying as a result of they characteristic free-flowing melodies and modal buildings. Extreme sample studying may end up in music that’s overly much like current music. The truth that Gregorian chants aren’t merely music however are additionally (perhaps primarily) an act of prayer, additionally contributes to the problem.
Coaching AI: token-based LSTM community © Regensburger Domspatzen
So how did they do it? They first ready a dataset of about 150 chants, later increasing it to over 9,000 melodies from GregoBase, an educational and copyright-free database of medieval notation. Based mostly on this dataset, they generated tokens that symbolize musical phrases and skilled their AI utilizing a Lengthy Brief-Time period Reminiscence (LSTM) mannequin. The AI’s aim was to foretell sequences in teams of forty tokens and generate new monophonic melodies. Sure parameters may very well be managed, such because the diploma of melismatics (what number of notes are used to brighten a single syllable) and the diploma of creativity. The latter was completed by incorporating a bigram rule into the algorithm. Bigrams are attribute sequences of two consecutive notes that replicate typical melodic progressions in a chant. The frequency with which explicit bigrams are employed can be utilized to explain the diploma of conventionality or creativity in a melody.
By adjusting how strongly the AI takes these bigram patterns under consideration when composing melodies, the person can management the extent of creativity. This constitutes probably the most technically subtle post-processing rule. Additionally they carried out musical guidelines, equivalent to forcing the ultimate notice to be the identical because the opening notice and guaranteeing {that a} remaining cadence is added to every chant. This remaining cadence guides the melody towards the ultimate notice (finalis) as a way to create a way of harmonic closure (for instance, in a Dorian chant, the sequence f–e–d). To attain this, appropriate endings had been extracted from the coaching information.
An instance of how neumes are transcribed in GABC © Regensburger Domspatzen
In a post-processing step, they utilized standards (music-theoretical and modal) to make sure that the ensuing melodies adhered to the principles of Gregorian chants. The AI system they developed can generate chants based mostly on GABC, a sort of programming language for Gregorian notation. Texts and neumes are entered as plain textual content after which transformed by particular software program into classical sq. notation (An apart: neumes are a sort of musical notation discovered above medieval textual content to assist the chanter keep in mind the grouping of pitches for every syllable, reasonably than actual notes).
The scholars obtained the Viewers Award on the Bundeswettbewerb KI (German Federal Synthetic Intelligence Competitors) in November 2025. They carried out certainly one of their newly AI-generated hymns entitled “Loquentes vobis,” (Latin for ‘Chatting with you”) to which they added biblical textual content from Ephesians 5:19 (this New Testomony verse encourages folks to speak and worship by means of music, particularly by “talking to at least one one other with psalms, hymns, and non secular songs, singing and making music in your hearts.”) The music was chanted in Regensburg Cathedral and obtained extraordinarily constructive suggestions from famend consultants together with from the Vatican, equivalent to Father Prof. Dr. Robert Mehlhart OP (Director of the Pontifical Institute of Sacred Music).
The result’s exported as GABC after which rendered. © Regensburger Domspatzen
Andreas Haberl informed me concerning the subsequent steps for the mission:
We’re at present summarizing the mission in a scientific paper. On the identical time, we plan to enter the mission into “Jugend forscht” (Youth Analysis), a Germany-wide competitors for younger researchers. We’re additionally making use of to take part within the Ars Electronica Pageant in Linz (Austria) this September. From a scientific perspective, our subsequent aim, after finishing the competitions, is to launch a publicly accessible interface for producing Gregorian chants. As well as, we wish to evaluate the efficiency of our present LSTM mannequin with different AI architectures, equivalent to transformers.
Though informal use of AI by the general public has grow to be virtually commonplace over the past couple years, researchers proceed to make advances in AI design. It actually impressed me {that a} group of highschool college students had been in a position to accomplish what they did. At RIKEN, we’ve a complete heart dedicated to AI analysis (RIKEN Heart for Superior Intelligence Challenge), and it even has a Music Info Intelligence Staff. It appeared apparent to ask them what they considered the mission. Postdoctoral researcher Shintaro Seki was type sufficient to provide me a number of feedback.
The scholars by the Danube River in Regensburg. © Regensburger Domspatzen
“I felt that the general design of the mission was extraordinarily nicely thought out, and that it was a mission made doable exactly as a result of it represents a collaboration between a traditionally established choir and pc science. By entrusting the interpretation of the rating and the manufacturing of sound to a choir, who’re deeply versed in church music, and limiting the function of the pc to a mechanism for acquiring musically coherent scores, the mission adopts a transparent division of labor.”
“This not solely defines in a compact means the issues that needs to be addressed by computation, but in addition succeeds in preemptively avoiding potential criticism or controversy that may come up when approaching spiritual matters by means of using computer systems, by preserving the act of choral singing—an exercise that may itself be thought of spiritual—inside the course of.”
Now, let’s take heed to it!
Concerning the Regensburger Domspatzen. In 975 CE, Bishop Wolfgang based his personal cathedral college, which, along with normal training, positioned explicit emphasis on musical coaching. This marked the delivery of the Regensburger Domspatzen. With their 1050-year custom, they’re in all probability one of many oldest boys’ choirs on this planet. Even again then, the scholars had been chargeable for liturgical singing within the bishop’s church. To this present day, they’re the cathedral choir of Regensburg Cathedral. They are often heard in St. Peter’s on Holy days, Sundays, and public holidays through the college 12 months.
