| # LinguaWave — Competition Writeup |
|
|
| **Task:** Language Identification — 8-class audio classification |
| **Metric:** Macro F1-score |
| **Dataset:** 8,000 train samples, 4,000 test samples (1,000 / 500 per language) |
| **Source:** Google FLEURS (CC BY 4.0) |
| **Languages:** Indonesian (id), Malay (ms), Vietnamese (vi), Thai (th), English (en), Chinese (zh), Arabic (ar), French (fr) |
|
|
| --- |
|
|
| ## Approach Progression |
|
|
| | # | Approach | Val Macro F1 | Description | |
| |---|----------|-------------|-------------| |
| | 1 | MFCC + SVM | **0.9775** | 40-dim MFCC mean+std → SVM RBF, C=10 | |
| | 2 | Pitch + LightGBM | **0.9594** | ~80-dim acoustic + prosody → LGBM | |
| | 3 | Bag of Codewords | **0.6290** | k-means codebook (k=64) + histogram → LogReg | |
| | 4 | CNN on Log-Mel | **0.8960** | 128×625 log-mel → 4-block CNN, BATCH=64, 5 epochs | |
| | 5 | Multi-Scale CNN | **0.9682** | 3-branch CNN (n_fft=512/1024/2048) + hard-neg mining on id/ms, precomputed mel cache | |
| |
| ### Key Takeaways |
| |
| - **Language ID is surprisingly easy with MFCCs**: at 0.9775, approach 1 already nears the ceiling. The 8 languages span distinct phoneme inventories, rhythm, and prosody, making them linearly separable in MFCC space. |
| - **More features ≠ better (approach 2 < 1)**: adding pitch and prosody (F0, jitter, shimmer, HNR) slightly *hurts* because these features are noisier over 10-second clips with variable speech content. The MFCC SVM is more robust. |
| - **Codebook approach underperforms (0.63)**: bag-of-codewords with k=64 compresses too aggressively. Language ID benefits from temporal dynamics (captured by delta-MFCC in approach 1) that flat histograms discard. |
| - **Hard negative mining for id/ms**: Indonesian and Malay share ~60% vocabulary and acoustic similarity. The multi-scale CNN addresses this by oversampling these pairs during training (2× factor). |
| - **Multi-scale spectrograms**: using n_fft=512/1024/2048 captures short-term phonemes (512), syllable-level rhythm (1024), and prosodic contours (2048) simultaneously. However, the multi-branch architecture converges unstably (F1 oscillating 0.30–0.78) — a lower learning rate and warm-up schedule are recommended. |
|
|
| --- |
|
|
| ## Dataset Details |
|
|
| - All clips from FLEURS dataset, filtered to 3–12 seconds duration |
| - Padded/trimmed to exactly 10 seconds (160,000 samples at 16 kHz) |
| - Stratified 80/20 split: 1,000 train / 500 test per language |
| - Class balance is exact — macro F1 = accuracy in this case |
|
|
| --- |
|
|
| ## Tips for Participants |
|
|
| 1. **Approach 1 sets a very high bar**: if your CNN underperforms the MFCC SVM (0.9775), your model is not learning — check preprocessing and batch size. |
| 2. **Memory-aware mel computation**: the mel shape is (1, 128, 625) at 16 kHz / 10s. With `BATCH=64` this is ~64 MB/batch — safe on Colab T4. Do NOT use batch >128 without checking memory. |
| 3. **id vs ms is the hard pair**: confusion between Indonesian and Malay is the main error source. Consider adding language pair augmentation or a dedicated hard-negative loss. |
| 4. **Cache your mel spectrograms**: approach 5 precomputes all mels to numpy arrays before training. This cuts training time from 27 hours (on-the-fly) to ~30 minutes. |
| 5. **LabelEncoder order matters**: ensure consistent label encoding between train/val/test splits. Always fit the encoder on training data only. |
| 6. **For approach 3 improvement**: try k=256 codebook, Fisher vectors instead of BoW histograms, and SVM classifier — this should recover 10–15 points vs. k=64 + LogReg. |
|
|
| --- |
|
|
| ## Hardest Language Pairs (id/ms) |
|
|
| Indonesian and Malay are the most commonly confused pair. Strategies that help: |
| - Train on more diverse speakers per language |
| - Use character n-gram language models on ASR transcripts |
| - Temporal pooling over the full 10-second clip rather than just mean/std |
|
|
| --- |
|
|
| ## Reproducibility |
|
|
| All scripts use `SEED = 42`. Feature extraction caches are keyed by script number (e.g., `lw_01_train.npy`). The CNN uses `torch.manual_seed(42)`. joblib parallel feature extraction is thread-based (prefer="threads") for librosa compatibility. |
|
|
| --- |
|
|
| ## Baseline Comparison |
|
|
| | Method | Val Macro F1 | Notes | |
| |--------|-------------|-------| |
| | Majority class | 0.125 | 8 balanced classes | |
| | Approach 3 (BoW k=64) | 0.6290 | Histogram loses temporal dynamics | |
| | Approach 4 (CNN Log-Mel) | 0.8960 | 5 epochs; converged well at epoch 5 | |
| | Approach 5 (MultiScale CNN) | 0.9682 | Best epoch 5; precomputed mel cache | |
| | Approach 2 (Pitch LGBM) | 0.9594 | Prosody features + gradient boosting | |
| | Approach 1 (MFCC SVM) | 0.9775 | Best overall — classical ML wins here | |
| | Human (expert) | ~0.98 | Native speakers, clean recordings | |
|
|
