Show HN: AI-generated music box version of any song (for baby)
A personal project that generates baby-friendly, slowed-down versions of adult music Uses BS-roformer for audio source separation (stemming) to isolate individual tracks Leverages Spotify's Basic Pitch to transcribe audio to MIDI for pitch/tempo adjustment The output has a "drunk music box" quality but is functional for the creator's use case Still in early development with acknowledged room for improvement
Analysis
TL;DR
- A personal project that generates baby-friendly, slowed-down versions of adult music
- Uses BS-roformer for audio source separation (stemming) to isolate individual tracks
- Leverages Spotify's Basic Pitch to transcribe audio to MIDI for pitch/tempo adjustment
- The output has a "drunk music box" quality but is functional for the creator's use case
- Still in early development with acknowledged room for improvement
Why It Matters
This project demonstrates a practical, creative application of open-source audio AI models (BS-roformer and Basic Pitch) to solve a real-world problem. It highlights how accessible music AI tools have become, enabling non-industry developers to build novel audio processing pipelines with existing components.
Technical Details
- BS-roformer: Used as the audio stemming component to separate mixed audio into individual instrument/vocal tracks, allowing selective processing of song elements
- Spotify Basic Pitch: Employed for audio-to-MIDI transcription, enabling tempo reduction and pitch shifting to create baby-friendly versions
- Pipeline architecture: Input audio → stemming (BS-roformer) → MIDI transcription (Basic Pitch) → tempo/pitch adjustment → output
- Output quality: Described as resembling a "drunk little music box" version, indicating artifacts from the multi-stage processing pipeline
- Implementation: Likely a Python-based pipeline combining these two open-source models
Industry Insight
- The growing ecosystem of specialized audio AI models (stemming, transcription, synthesis) enables rapid prototyping of niche audio applications without building models from scratch
- Consumer-facing AI audio tools are moving beyond technical demos into personal utility, suggesting market opportunities in specialized audio processing for underserved audiences
- The "drunk music box" quality underscores the current limitations of multi-stage audio processing pipelines and the gap between research demos and production-ready quality
Disclaimer: The above content is generated by AI and is for reference only.