HyVoxPopuli Dataset

The HyVoxPopuli dataset is the Armenian language subset of the Facebook VoxPopuli dataset. The name "HyVoxPopuli" comes from combining "hy" (the ISO 639-1 language code for Armenian) with "VoxPopuli" (the original dataset name). It is a high-quality collection of Armenian speech recordings with expert-validated transcriptions, carefully extracted and processed from the original VoxPopuli dataset (https://github.com/facebookresearch/voxpopuli).

This dataset is specifically designed for training and evaluating Armenian Automatic Speech Recognition (ASR) models, making Armenian speech technology more accessible to researchers and developers.

Dataset Structure

The dataset is provided in Parquet format with the following splits:

• Train: 498 examples
• Validation: 62 examples
• Test: 63 examples

Features

{
    'audio_id': string,       # Unique identifier for the audio segment
    'audio': {                # Audio data
        'path': string,       # Path to audio file
        'array': float32[],   # Audio samples
        'sampling_rate': int  # 16000 Hz
    },
    'raw_text': string,       # Original transcription
    'normalized_text': string, # Normalized transcription
    'gender': string,         # Speaker gender
    'speaker_id': string,     # Unique speaker identifier
    'is_gold_transcript': bool, # Expert validation flag
    'accent': string         # Accent information
}

Loading the Dataset

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("Edmon02/hyvoxpopuli")

# Load a specific split
train_dataset = load_dataset("Edmon02/hyvoxpopuli", split="train")

Usage Examples

Here's a simple example of how to use the dataset for training an ASR model:

import torch
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

# Load the dataset
dataset = load_dataset("Edmon02/hyvoxpopuli")

# Load pre-trained model and processor
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-large-xlsr-53")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-large-xlsr-53")

# Prepare a batch
batch = dataset["train"][0]
input_values = processor(batch["audio"]["array"], 
                       sampling_rate=16000, 
                       return_tensors="pt").input_values

# Get model predictions
logits = model(input_values).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.decode(predicted_ids[0])
print(f"Predicted text: {transcription}")
print(f"Original text: {batch['normalized_text']}")

Dataset Creation

The dataset was created by:

1. Extracting Armenian language recordings from VoxPopuli
2. Filtering for high-quality audio segments
3. Expert validation of transcriptions
4. Converting to a standardized format with normalized text

License

This dataset is licensed under CC-BY-4.0.

Citation

@dataset{hyvoxpopuli2023,
    title = {HyVoxPopuli: Armenian Speech Recognition Dataset},
    year = {2023},
    publisher = {Hugging Face},
    journal = {Hugging Face Datasets},
    url = {https://huggingface.co/datasets/Edmon02/hyvoxpopuli}
}