Choir Blend Optimization With Phone Audio Analysis

ISEF Category: Technology Enhances the Arts

Subcategory: Music and Image Manipulation  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A choir can sound like one voice even when many singers are on stage. Small timing and tone differences decide whether the sound fuses or feels messy. Your project can measure that blur between individual voices and shared blend. Then you can test what changes make the ensemble sound more unified.

What Is It?

This project studies choir blend, which means how well separate voices merge into one smooth sound. Think of it like mixing paint. If the colors stay separate, you see each one. If they blend well, you get a single shade. In music, blend depends on pitch, timing, loudness, vowel shape, and how each singer’s sound fills the room.

You can record a small ensemble with several phone microphones, then separate the voices with a source-separation tool such as Demucs-vocal. That gives you a way to inspect each singer’s signal on its own. You can then look at things like timing offset, loudness balance, and formants, which are the resonance bands that help define vowel sound. A learned perceptual fusion metric means a model that predicts how fused the choir sounds to listeners, based on human rating data you collect.

Why This Is a Good Topic

This topic works well because you can measure real audio data, compare multiple singing conditions, and connect the numbers to human perception. It links computer science, signal analysis, and music. You can test a clear question, like which adjustments improve blend, and you can score the results with listener ratings and audio features. That makes the project both technical and artistic.

Research Questions

• How does small timing correction between singers change perceived choir blend?
• What is the effect of per-singer EQ adjustment on listener-rated fusion?
• Does aligning vowel formants improve measured blend more than matching loudness alone?
• To what extent does source-separated voice analysis predict human blend ratings?
• Which combination of timing, EQ, and balance produces the highest perceptual fusion score?

Basic Materials

• Smartphone or tablet microphones for multitrack recording.
• Quiet room with controlled acoustics.
• Small ensemble of singers.
• Laptop computer for audio analysis.
• Headphones for playback checks.
• Spreadsheet software for rating and analysis.
• Consent forms for all participants.

Advanced Materials

• Multi-channel audio interface.
• Matched condenser microphones.
• Acoustic treatment panels.
• Reference speaker for calibration playback.
• Laptop or workstation with GPU support.
• Dataset storage drive.
• Professional audio editing software if available, plus Python environment for analysis.

Software & Tools

• Demucs: Separates mixed vocal recordings into individual voice stems for analysis.
• Python: Runs audio feature extraction, model fitting, and statistical tests.
• Librosa: Extracts timing, pitch, and spectral features from recordings.
• ImageJ: Can help inspect spectrogram images if you export them for comparison.
• Google Forms: Collects listener blend ratings in a simple, organized way.

Experiment Steps

1. Define the exact blend question you want to test, such as timing, EQ, or both.
2. Choose one recording setup and keep it fixed so your measurements stay comparable.
3. Plan a listener-rating method that turns subjective blend into numeric data.
4. Build a feature set from the audio, such as timing offsets, spectral balance, and formant alignment.
5. Fit a model that links those audio features to listener ratings, then check how well it predicts new samples.
6. Compare different adjustment strategies and decide which one improves both the measured signal and the human rating.

Common Pitfalls

• Recording each singer with a slightly different mic distance, which makes loudness comparisons unreliable.
• Letting room echo change between sessions, which can hide real blend differences.
• Using too few listener ratings, which makes the perceptual fusion score noisy and unstable.
• Treating separated voice stems as perfect ground truth, which can distort timing and EQ analysis.
• Changing several variables at once, which makes it hard to tell whether timing, EQ, or balance caused the result.

What Makes This Competitive

A strong version of this project goes past simple before-and-after comparisons. You would need careful controls, a clear listener study, and a model that predicts blend from audio features rather than guessing from one metric. The best entries will test whether perceived fusion matches measurable timing and spectral alignment across different singers and songs. That gives you a real analytical contribution, not just a demo.

Project Variations

• Test whether vowel matching matters more than loudness matching for perceived blend.
• Compare beginner choirs, advanced choirs, and solo-trained singers to see which group produces the strongest fusion.
• Use only mobile phone recordings, then test how much blend prediction drops compared with studio-grade microphones.

Learn More

• PubMed: Search for review articles on auditory perception, ensemble singing, and timbre fusion.
• NIH PubMed Central: Read free full-text papers on audio perception and machine learning methods.
• MIT OpenCourseWare: Look for free courses on signal processing and machine learning basics.
• NOAA Sound and Acoustic Research resources: Find public educational material on how sound behaves in real environments.
• The Journal of the Acoustical Society of America: Search articles on choir blend, vocal timbre, and auditory fusion through your school library or publisher previews.