MusicGen_Lyre_0_New

Author: Richard Jiang

Overview

This fine‑tuned MusicGen model generates Ancient Greek lyre instrumentals with a structured musical form. The model was trained on a curated dataset of several hundred lyre tracks, each annotated for mood, tempo, tonality—and also paired with a global prompt emphasising the desired structure: melody → variation → return. The result is a generative system that prioritises musical form over surface detail, aiming to capture the essence of traditional lyre compositions.

Background & Architecture

This model is built on Meta’s MusicGen architecture—a single-stage autoregressive transformer trained over a 32 kHz EnCodec tokenizer using four codebooks at ~50 Hz. MusicGen facilitates controllable music generation from textual prompts, achieving high-quality outputs efficiently (Replicate, Hugging Face).

Intended Use

• Project Showcase / Research Prototype: Ideal for demonstrating structured generation of Ancient Greek lyre-style music.
• Educational & Cultural Preservation: Useful for researchers exploring AI-driven preservation and regeneration of traditional music.

Not intended for commercial distribution under current license constraints.

Usage Instructions

YYou can run this model via the Replicate API:

import replicate

output = replicate.run(
  "richardjiang736/musicgen_lyre_0_new:latest",
  input={
    "prompt": "medieval-flavoured lyre melody in Aeolian mode, clear motif then variation then return",
    "duration": 30,
    "classifier_free_guidance": 3.5,
    "temperature": 1.0,
    "top_k": 250,
    "top_p": 0.0
  }
)
print(output)

Parameters:

• duration: Typically 30 seconds; can be extended via continuation logic.
• temperature, top_k, top_p: Standard sampling controls for diversity versus structure.

Model Evaluation & Analysis

• Baseline Output: Early results show recognizable motif → variation → return structure, with room for tonal refinement.
• Structural coherence: Does the generated track follow the three-part arrangement?
• Stylistic fidelity: Does the timbre approximate lyre-like qualities?

• Iteration Strategy:

• TTest with explicit prompts vs. style-only prompts to measure reliance on one_same_description. - SSweep CFG values to balance creativity and prompt adherence. - CComparative listening across modal changes (Aeolian, Dorian, etc.)

Ethical & Cultural Considerations

• Cultural Respect: This model is trained on publicly available Ancient Greek lyre recordings and annotations. While generative in nature, it does not claim authenticity or replicate living tradition.
• Access & Democratization: The model aims to make traditional musical forms accessible and regenerable, not to replace human performance or scholarship.
• Licensing: Built upon MusicGen weights licensed under CC‑BY‑NC. Use of outputs may be restricted for non-commercial purposes(news.ycombinator.com, Replicate, Hugging Face)
• Limitations: This is a proof-of-concept tool—user caution advised when interpreting generated content as historically accurate.

References

• CCopet et al., Simple and Controllable Music Generation (AudioCraft / MusicGen architecture)(arXiv)
• MMeta’s MusicGen README on Replicate platform(Replicate)
• AAudioCraft / Hugging Face documentation on MusicGen usage and architecture(Hugging Face)

Model Card

Future Directions

• Expand dataset with longer and more varied lyre recordings.
• Integrate audio as conditioning seed to test continuation-based prompting.
• Collaborate with ethnomusicologists to fine-tune model’s stylistic authenticity.
• Develop interactive demos comparing human-performed lyre tracks vs. generated outputs.

Contact & Citation

To cite or collaborate:

Richard Jiang Replicate Model Page

Email: richardjiang736@gmail.com

If you publish work based on this model, a citation would be appreciated.

Thank you for exploring MusicGen_Lyre_0_New. This is an experimental step toward AI‑powered cultural preservation, with structure baked in—and variation inspired.