EchoScript: script-aware music generation

BSc (Hons) Computer Science (Artificial Intelligence) dissertation, Heriot-Watt University, Dubai. Supervisor: Kayvan Karim. Graded First Class.

Question

Generative music models can produce plausible audio, but soundtrack work needs more: melodic coherence, thematic continuity, and music that follows the emotional arc of a scene. EchoScript asks whether conditioning on the film script itself can close that gap.

Method

• Fine-tuned MusicGen, a transformer audio model, on a curated dataset annotated with genre, mood, instrumentation, tempo and narrative context.
• Conditioned soundtrack generation on film scripts, so that the music follows the story rather than a short text prompt.

Findings

• Objective: lower Fréchet Audio Distance than the base model (4.37 vs. 4.55).
• Human: in structured listening tests, participants preferred the fine-tuned model for both musical quality and narrative alignment.

Using both an automatic metric and listeners, and seeing where they agree, is the thread that runs from this project into my later work on evaluation.

Published as EchoScript: Enhancing AI Music Generation for Cinematic Scoring via Script-Aware Fine-Tuning, Proceedings of the AAAI Symposium Series 6(1), 2025.

Limitations and next steps

The FAD comparison and listening preferences address different aspects of generation quality. They do not establish that a model meets every requirement of professional scoring. The paper discusses technical constraints and future work on symbolic-audio integration, audio mixing and standardized evaluation metrics.

Sources

Read the published paper or the complete abstract and citation. The waveform is decorative; the two bar values remain fixed at the reported rounded FAD scores.