UPS Challenge Baselines

Reference implementation for loading data from the Unsupervised People's Speech (UPS) Challenge dataset.

Overview

This repository provides a WebDataset-based dataloader for efficiently streaming and processing audio data from the UPS dataset. It includes support for language-specific filtering using soft language ID markers.

Structure

ups_challenge/
├── dataloaders/
│   ├── base.py           # Core dataloader and preprocessing
│   ├── urls.py           # URL building and language filtering
│   └── build_index.py    # Language ID index construction
└── examples/
    └── main.py           # Usage example with wav2vec2

Features

• Streaming data loading via WebDataset from HuggingFace
• Language-specific filtering using pre-computed language ID indices
• Audio preprocessing: resampling, mono conversion, chunking
• Random chunk sampling for long audio files
• Automatic fallback: loads first 4 tar files if no languages specified

Usage

Basic Example

from ups_challenge.dataloaders.base import build_wds_dataset, collate_fn
import torch

# Load specific languages (e.g., English)
wds_dataset = build_wds_dataset(langs=['en'])

# Or load all languages (uses first 4 tar files by default)
wds_dataset = build_wds_dataset(langs=[])

data_loader = torch.utils.data.DataLoader(
    wds_dataset,
    batch_size=1,
    num_workers=1,
    collate_fn=collate_fn,
)

for batch in data_loader:
    if batch is None:
        continue
    # batch contains 'input_values' and 'attention_mask'
    print(batch['input_values'].shape)

Running the Example

From the project root:

python -m ups_challenge.examples.main

Requirements

• torch
• transformers
• webdataset
• torchcodec
• braceexpand
• tqdm

Set your HuggingFace token:

export HF_TOKEN=your_token_here

Language Filtering

The dataloader supports filtering by language codes (e.g., 'en', 'es', 'fr'). When languages are specified:

1. A language ID index is built (or loaded from ./data/lid_index.pkl)
2. Only tar files containing the specified languages are streamed
3. This significantly reduces download time and storage requirements

Without language specification, the loader defaults to the first 4 tar files for quick testing.

Audio Processing

• Target sample rate: 16kHz
• Chunk duration: 10 seconds
• Max chunks per file: 16 (for long files)
• Output format: [N_chunks, samples] tensor with attention mask

Short files (<10s) are padded to chunk length; long files are randomly sampled.