Hellenistic Central Asia through the Eyes of GenAI – Part 2: Music

This is part two of a three-part series on the biases about Hellenistic Central Asia in generative artificial intelligence (AI) datasets. Last time, I discussed how generative image AI tools were creating images of Hellenistic Central Asia that were conglomerations of the available sources in its dataset. Due to the scarcity of ancient sources for Hellenistic Central Asia, these datasets are mostly populated by modern representations of these ancient people, and so the outputs including them are biased towards these modern presentations. This scarcity is even more pronounced in AI-generated music.

Since ChatGPT’s surge of popularity in November 2022, some theorize that generative AI tools could be the answer to uncovering how ancient music may have actually sounded. Although this seems like an exciting prospect, there are several factors surrounding the current capacities of generative music AI tools and the available sources for ancient music that must be considered first.

Generative music AI tools produce a wide variety of sound outputs that vary in length, style, and complexity. These tools use a combination of semantic (text) data and acoustic (audio) data to identify the requested output from the text prompt, and then generate a relevant audio clip based on its training dataset. Some models also include image data to create cover art for the generated music as well. The actual content of these datasets have not yet been publicly released, much like for most generative AI tools on the market today. However, one thing is clear: there are no audio clips of ancient music in any of these datasets.

The earliest known audio recording comes from 1860, almost two millennia since Hellenistic Central Asia. Unfortunately, this also means that there are no audio recordings of ancient music. The most that we have are the brilliant reconstructions of ancient instruments made and played by archaeologists and musicians like Stefan Hagel and Bettina Joy de Guzman. As such, the source base for AI-generated music is entirely skewed towards modern understandings of ancient music.

To see what kind of audio was produced by these tools at their current state, I tested out two models, CassetteAI and Suno. While both are forms of the diffusion model, machine learning algorithms that degrade and reconstruct their training datasets to enhance generated outputs, each creates a different type of output: CassetteAI focuses more on instrumentals, while Suno produces lyrics, vocals, backing music, and cover art all at once. I prompted each tool twice to see what features of Hellenistic Central Asia and Central Asian music would appear in the outputs, once including Bactria and the other including Ai Khanoum, a Hellenistic city in Bactria.

CassetteAI produced short, instrumental beats in both cases. When prompted with, “Write a song about Ai Khanoum in the style of Central Asian music,” the tool outputted a jaunty tune that had heavy drum beats, a solo flute, and some kind of electronic string instrument.

Figure 1: Cassette, CassetteAI 3.0.1, 10 March 2024 version, personal communication, generated on 14 January 2025 by Edward A. S. Ross. Prompt: “Write a song about Ai Khanoum in the style of Central Asian music.”

Drums, pipes, and string instruments are quite common in musical traditions across Central Asia today. Recordings of several styles of Central Asian folk music are available through YouTube and the Smithsonian Folkways Recordings series, so it is possible that these CassetteAI generations were drawing on these details in its training data. Bearing that in mind, this CassetteAI generation sounds like a backing track from a vaguely desert-like level in a video game rather than a traditional folk song.

My second attempt with CassetteAI used the prompt, “Write a song in the style of Hellenistic Central Asian music about life in Bactria,” in order to draw out conceptions of ancient music in the Central Asian context.

Figure 2: Cassette, CassetteAI 3.0.1, 10 March 2024 version, personal communication, generated on 14 January 2025 by Edward A. S. Ross. Prompt: “Write a song in the style of Hellenistic Central Asian music about life in Bactria.”

This output features a lot more metallic sounds, such as gongs and bells, alongside the drums, plucked strings, and flutes. It also feels like it is reminiscent of a video game backing track for a Middle Eastern bazaar, such as the Marketplace from Assassin’s Creed: Mirage (2023).

Suno, on the other hand, produced outputs that had instrumentals, lyrics, vocals, and cover art, but the mystical bias is much more overt in these features. Using the same prompts as above, Suno created two outputs with the same lyrics but different musical styles for each input. For the prompt, “Write a song about Ai Khanoum in the style of Central Asian music” it produced:

Figure 3: Suno Inc., Suno v4, 19 November 2024 version, personal communication, generated on 14 January 2025 by Edward A. S. Ross. Prompt: “Write a song about Ai Khanoum in the style of Central Asian music.”

and:

Figure 4: Suno Inc., Suno v4, 19 November 2024 version, personal communication, generated on 14 January 2025 by Edward A. S. Ross. Prompt: “Write a song about Ai Khanoum in the style of Central Asian music.”

These songs are much more ethereal and appear to include vocal oscillations reminiscent of North Indian vocal ornamentation (alankāra). The cover art reinforces this mystical idea by primarily featuring sand dunes and distant mountains. The lyrics describe Ai Khanoum minimally, but its main characteristic here is trade. The references to “marbled halls and golden dreams,” “gem-laden paths of spices gold,” and “history’s echo in the bazaar” appear to reference different time periods of the Silk Road all at once. These songs are quite orientalist at their core, but that is likely due to the orientalist bias present in most 20^th-century scholarly sources about Hellenistic Central Asia.

The second prompt produced similar cover art, featuring sand dunes, but also include (Bactrian?) camels and distant walled cities. However, the musical genre was more europop or rock.

Figure 5: Suno Inc., Suno v4, 19 November 2024 version, personal communication, generated on 14 January 2025 by Edward A. S. Ross. Prompt: “Write a song in the style of Hellenistic Central Asian music about life in Bactria.”

Figure 6: Suno Inc., Suno v4, 19 November 2024 version, personal communication, generated on 14 January 2025 by Edward A. S. Ross. Prompt: “Write a song in the style of Hellenistic Central Asian music about life in Bactria.”

These songs are quite fun, but they are in no way reminiscent of ancient music. The lyrics do appear to include features connected to Bactria and Hellenistic Central Asia. “Sips of wine” may refer to the popularity of wine drinking in Hellenic cultures, “the temple’s bell” may refer to the “Temple with Indented Niches” uncovered at Ai Khanoum, and “in Bactria’s beauty we climb” possibly referring to the region’s close proximity to the Pamirs and Hindukush.

Looking at the current state of generative music AI tools, it is quite clear that we are far away from generative AI tools producing ancient music. When it comes to Hellenistic Central Asia, we tend to see the distillation of all the digitally available sources discussing it, and the average result is heavily biased by modern Central Asian musical traditions or the orientalist perceptions about the region in Western scholarship and media. Although these AI-generated songs can be quite catchy and human-sounding, we need to be continually vigilant for biases surfacing in the outputs to prevent misunderstanding and perpetuating misconceptions.

In the next part of this series, I will explore the biases about Hellenistic Central Asia present in generative AI text tools.

---

Cover Image: Suno Inc., Suno v4, 19 November 2024 version, personal communication, generated on 14 January 2025 by Edward A. S. Ross. Prompt: “Write a song in the style of Hellenistic Central Asian music about life in Bactria.”

---

References

Agostinelli, Andrea, Timo I. Denk, Zalán Borsos, Jesse Engel, Mauro Verzetti, Antoine Caillon, Qingqing Huang, Aren Jansen, Adam Roberts, Marco Tagliasacchi, Matt Sharifi, Neil Zeghidour, and Christian Frank. “MusicLM: Generating Music From Text.” arXiv. 2023. Accessed 20 January 2025. https://arxiv.org/pdf/2301.11325.

Angelides, Brendan. Marketplace [Song]. Ubisoft Music. 2023.

Cain, David. “Reviving Ancient Music through AI.” Texploration with David Cain. LinkedIn, November 23, 2023. https://www.linkedin.com/pulse/unearthing-harmonies-journey-revival-ancient-music-through-david-cain-csfhc/.

Cassette. CassetteAI 3.0.1 (10 March 2024 version) [Text-to-sound model]. 2023. https://cassetteai.com/.

de Guzman, Bettina Joy. bettinajoydeguzman1981’s Channel [YouTube page]. YouTube. Accessed January 17, 2025. https://www.youtube.com/@bettinajoydeguzman1981.

Edison was not the first to record the human voice…. BBC Radio 4 in Four. Accessed January 17, 2025. https://www.bbc.co.uk/programmes/articles/4zzbCwh39KR33grNt6sXxMz/edison-was-not-the-first-to-record-the-human-voice.

Francfort, Henri-Paul. “Ai Khanoum ‘temple à niches indentées’ (temple with indented niches) and Takht-i Sangin ‘Oxus temple’ in historical cultural perspective: a hypothesis about the cults.” Parthica 14 (2012): 109-136.

Hagel, Stefan. Profile [Academia.edu page]. Academia.edu. Accessed January 17, 2025. https://oeaw.academia.edu/StefanHagel.

Levin, Theodore, Saida Daukeyeva, and Elmira Köchümkulova. The Music of Central Asia. Bloomington: Indiana University Press, 2016.

Ross, Edward A. S. “Hellenistic Central Asia through the Eyes of GenAI – Part 1: Images.” The Digital Orientalist, October 22, 2024. Accessed January 15, 2025. https://digitalorientalist.com/2024/10/22/hellenistic-central-asia-through-the-eyes-of-genai-part-1-images/.

Rowell, Lewis. Music and Musical Thought in Early India. Chicago: University of Chicago Press, 2015.

Smithsonian. “Music of Central Asia.” Smithsonian Folkways Recordings. Accessed January 16, 2025. https://folkways.si.edu/music-of-central-asia-series.

Suno Inc. Suno v4 (19 November 2024 version) [Text-to-sound model]. 2023. https://suno.com/.