File size: 1,930 Bytes
233b2e6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
# V31 Feature Engineering Notes

## Why we did V31
V16 is stable and strong, but it may saturate. V31 adds richer signal descriptors to capture information that plain MFCC/chroma stats can miss.

Goal:
- keep V16 baseline safety,
- add meaningful audio dynamics/rhythm/texture features,
- compare objectively against V16 before upload.

## What was added in V31
1. **MFCC Delta + Delta-Delta**
   - Captures temporal changes in timbre.
   - Useful when genres differ by evolving patterns, not just static spectral shape.

2. **HPSS (Harmonic-Percussive Source Separation)**
   - Splits audio into harmonic and percussive components.
   - Helps distinguish melody-heavy vs beat-heavy genre characteristics.

3. **Tempo / Beat Strength Features**
   - Adds rhythmic cues (tempo, onset strength mean/std).
   - Useful for genre rhythm identity.

## Where implemented
- Notebook: `notebooks/DL_21f2000735_FINAL_UPLOAD_V31.ipynb`
- Demo feature output script: `src/v31_feature_demo_output.py`
- Demo output file: `reports/v31_demo_feature_output.json`

## How to use it in decisions
- Run V31 compare block (V16-base vs V31-fusion).
- Check:
  - `v16_base_macro_f1`
  - `v31_fusion_macro_f1`
  - `delta_v31_minus_v16`
- Upload V31 only if delta is non-negative and run is stable.

## How to learn more
1. **Librosa docs**
   - MFCC / delta / HPSS / tempo APIs
2. **Audio feature engineering guides**
   - spectral contrast, rhythm features, pooling strategies
3. **Ablation practice**
   - turn one feature family on/off and record F1 delta
4. **Error analysis loop**
   - inspect confusion matrix and add features targeting weak class pairs

## Practical next learning experiments
1. HPSS-only ablation vs base
2. Delta-only ablation vs base
3. Tempo-only ablation vs base
4. Combined (V31) vs V16 baseline
5. Class-wise F1 comparison for weak genres

This note exists to keep V31 explainable in viva and reproducible in future runs.