fassabilf/linguawave-competition

LinguaWave — Language Identification Competition

Pelatnas IOAI 2026 | Task 2 of 3

Identify the language of a 10-second speech clip from 8 languages. Compete to achieve the highest Macro F1-score on the test set.

Task

Input: .wav audio file (10 seconds, 16 kHz mono)
Output: Language code from {id, ms, vi, th, en, zh, ar, fr}
Metric: Macro F1-score

Languages

Code	Language	Region
id	Indonesian	Southeast Asia
ms	Malay	Southeast Asia
vi	Vietnamese	Southeast Asia
th	Thai	Southeast Asia
en	English	Global
zh	Chinese (Mandarin)	East Asia
ar	Arabic	Middle East
fr	French	Europe

Dataset

Split	Samples	Per-language
Train	8,000	1,000
Test	4,000	500

Source: Google FLEURS (CC BY 4.0)

File Structure

├── train/
│   ├── id/
│   │   └── *.wav
│   ├── ms/
│   │   └── *.wav
│   └── ...
├── test/
│   └── *.wav           # flat, unlabeled
├── train.csv           # id, label
├── test.csv            # id (no label)
├── sample_submission.csv
├── solution.csv        # ground truth
├── notebooks/          # 6 Colab-ready approaches
│   ├── 00_starter.ipynb
│   ├── 01_mfcc_svm.ipynb          # F1=0.9775
│   ├── 02_pitch_lgbm.ipynb        # F1=0.9594
│   ├── 03_bag_of_codewords.ipynb  # F1=0.6290
│   ├── 04_cnn_mel.ipynb           # F1~0.95
│   └── 05_multiscale_cnn.ipynb    # F1=0.9682
├── submissions/
└── writeup/
    └── writeup.md

How to Load

import pandas as pd, librosa
train_df = pd.read_csv("train.csv")
y, sr = librosa.load(f"train/{train_df.iloc[0]['id']}", sr=16000)
print(f"Language: {train_df.iloc[0]['label']}")  # e.g. 'id'

Notebooks (Colab-ready)

Notebook	Approach	Val Macro F1
01_mfcc_svm	MFCC + SVM RBF	0.9775
02_pitch_lgbm	Acoustic+Prosody + LightGBM	0.9594
03_bag_of_codewords	k-means codebook + LogReg	0.6290
04_cnn_mel	Log-mel (128×625) + CNN	0.8960
05_multiscale_cnn	3-branch CNN + hard-neg mining	0.9682

Tip: The hardest pair is Indonesian (id) vs Malay (ms) — they share ~60% vocabulary. Focus your improvements there.

Citation

FLEURS: Few-Shot Learning Evaluation of Universal Representations of Speech.
Conneau et al., SLT 2022.