# The Silence Inside

By Peter Howitt, 28 August 2025.
Web page: https://compossible.blog/2025/08/28/the-silence-inside/
Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).

---


Experimental AI album

I also tried my hand at an AI generated album '[The Silence Inside](<https://www.bandlab.com/poetones/albums/7688ead4-e183-f011-b482-000d3a97c7d6>)': a mixture of some of my poetry and instrumental music all using similar prompts to give the whole thing a (hopefully) coherent sound. Unfortunately AI can not yet produce studio quality tracks.

Listen With or Without Prejudice

![](/images/uploads/2025/11/the-silence-inside-album-cover.webp)[The Silence Inside](<https://www.bandlab.com/poetones/albums/7688ead4-e183-f011-b482-000d3a97c7d6>) ![](/images/uploads/2025/11/the-silence-inside.webp)_AI generated image made by author._

* * *

**Notes:**

AI music generators (like Suno.ai) leave a lot of artefacts (including shimmering and hissing). This is due to the way they create music. The process involves the following stages: 

  1. **Prompt Interpretation (NLP):** When a user enters a text description (e.g., "neo-classical piano with iceberg sounds") or custom lyrics. Suno uses a language model to understand the creative vision, mood, genre, and specific instructions.
  2. **Music Generation:** The interpreted text is used to condition a generative audio model, using a combination of transformer-based and diffusion techniques. 
     * **Data Training:** The AI model is trained on a massive dataset of music, learning patterns, chord structures, melodies, rhythms, and how these elements correlate with different descriptions and tags.
     * **Audio Generation:** The AI model generates the raw audio waveform from scratch each time effectively bringing sounds out of a sort of white noise. It essentially "learns" how instruments and vocals sound and should progress together.
  3. **Vocal Synthesis:** A specialised text-to-singing model is used to generate human-like vocals that align with the provided lyrics, timing, and style of the song. This model is adept at creating natural phrasing and emotional nuances.
  4. **Post-Processing:** The generated components (instrumentals and vocals) can then be mixed and "mastered" in applications like Bandlab (for amateurs like me) or with digital audio workstation (DAW) techniques.