Multilingual Yodas Streaming Dataset (OWSMv4 Optimized)

This repository provides a streaming-ready version of the Yodas (Yielding Optimized Data from Any Source) dataset. It is specifically designed for training and fine-tuning large-scale Multilingual ASR models (such as OWSM v4, Mamba-based architectures, and beyond) without the need to download terabytes of raw data.

By utilizing a custom Lhotse-based loading script, this dataset allows researchers to stream audio data directly into their training pipeline using curl and ffmpeg.

🚀 Key Features

Massively Multilingual: Designed to support up to 75 languages, providing a unified interface for global speech research.
Zero-Download Streaming: Access over 20 million audio cuts in real-time. No massive local storage is required.
Lhotse Integration: Uses Lhotse manifests for precise audio slicing and robust metadata management.

🛠️ Prerequisites

To ensure the streaming loader functions correctly, your environment must meet the following requirements:

1. System Dependencies

The loader relies on system-level pipes to stream audio. Ensure these are installed and added to your PATH:

FFmpeg: Required for real-time audio decoding.
curl: Required for streaming data from remote servers.

2. Python Libraries

Due to changes in the Hugging Face datasets library (v3.0+), you must use a version lower than 3.0.0 to support custom loading scripts.

pip install "datasets<3.0.0" lhotse torchaudio smart_open requests

💻 Usage

To use this dataset, you must enable trust_remote_code=True and streaming=True.

from datasets import load_dataset

# Load the dataset in streaming mode
dataset = load_dataset(
    "N02N9/yodas_owsmv4_streaming", 
    trust_remote_code=True, 
    streaming=True
)

# Accessing the data (IterableDataset)
train_iter = iter(dataset['train'])
sample = next(train_iter)

print(f"ID: {sample['id']}")
print(f"Language: {sample['id'].split('_')[-2]}") # Language code extraction
print(f"Text: {sample['text']}")
print(f"Audio Sample Rate: {sample['audio']['sampling_rate']}Hz")

📊 Dataset Features

Feature	Type	Description
id	string	Unique identifier for each audio cut (includes language info)
audio	dict	Real-time decoded audio (array, sampling_rate=16kHz)
text	string	The transcribed text for the audio segment

⚠️ Troubleshooting & FAQ

KeyError: 'text': This usually happens if the loader script (yodas_owsmv4_streaming.py) failed to run or if you are using a cached version of the dataset. Try adding download_mode="force_redownload" to the load_dataset function.
Hanging on Windows: As mentioned above, this loader is not compatible with Windows. Please use a Linux/WSL2 environment.
Library Version: If you get a "Loading script unsupported" error, downgrade your datasets library: pip install "datasets<3.0.0".

📜 License & Acknowledgments

Original Source: Yodas Dataset (Meta/ESPnet)
Modifications: Custom Lhotse-based streaming loader and offset corrections by N02N9.
License: This dataset follows the licensing terms of the original Yodas release (typically CC-BY 4.0). Please credit the original authors accordingly.