darvilab/nepali-asr-community-data

Nepali ASR Open Data Collection

Dataset Description

This dataset contains audio recordings of spoken Nepali, contributed by the community, aimed at advancing Automatic Speech Recognition (ASR) technology for the Nepali language. The data includes a variety of speakers, prompts, and natural conversational speech, including instances of Nepali-English code-switching.

The data is collected through our public Gradio application: Contribute Your Voice Here!

We encourage everyone to contribute and help us build a robust, open-source resource for Nepali speech technology.

Languages: Nepali (नेपाली), with common English code-switching.

License: Apache

How to Use

The dataset is structured with individual audio files and corresponding metadata entries.

Data Structure

Each data point typically consists of:

• audio: The audio file (usually .wav). The audio is loaded automatically by the datasets library.
• id: A unique identifier for the recording.
• text: The transcribed text that was spoken (either user-provided, a fixed prompt, or AI-generated).
• gender: Speaker's self-reported gender.
• age_group: Speaker's self-reported age group.
• ethnicity: Speaker's self-reported ethnicity.
• last_name: Speaker's self-reported last name (used for accent diversity analysis, may be generalized or anonymized in future versions).
• region: Speaker's self-reported geographical region within Nepal.
• emotion: Speaker's self-reported emotion during recording.
• recording_type: The source of the text spoken (e.g., "free_text", "prompted_text", "ai_generated_text").
• timestamp: The timestamp of the original recording submission.
• (Potentially other metadata like mother_tongue, recording_environment)
• Rating Information (upvotes, downvotes, etc.): Collected through community review, these may be in separate files (ratings_entries/{id}.json) or aggregated.

The audio files are located in the audio/ directory, individual metadata entries in metadata_entries/, and individual rating entries (if applicable) in ratings_entries/.

Using with datasets

from datasets import load_dataset

dataset_id = "[darvilab/nepali-asr-community-data]"

# To load based on individual metadata JSON files:
# This assumes your metadata_entries/*.json files also include a key like 'hf_audio_path'
# that points to the audio file within the repo, e.g., "audio/some_id.wav"
# The `audio` feature in dataset_info should be configured to cast this path.
# Alternatively, if your JSONs are structured to be directly consumable by `load_dataset("json", ...)`
# and then you map the audio:
try:
    # Attempt to load if metadata_entries contain sufficient info for direct loading
    # This might require your JSONs in metadata_entries to also include a reference to the audio file
    # that the 'audio' feature can process.
    # For example, if your metadata_entries/{id}.json has:
    # { "id": "...", "text": "...", "hf_audio_path": "audio/id.wav", ... }
    # And your dataset_info.features.audio is set up to interpret hf_audio_path.
    # This is a more advanced setup.

    # A simpler approach if you have an aggregated metadata file (see collection app's admin sync):
    # dataset = load_dataset(dataset_id, data_files="aggregated_metadata_with_audio_paths.jsonl")

    # For now, let's assume you'll manually iterate or provide a loading script.
    # The following is a placeholder for how one might load it if structured for direct loading.
    # This will likely need adjustment based on your final file structure on HF.
    print(f"To load this dataset, ensure your files in 'metadata_entries/' and 'audio/' are structured correctly "
          f"or use a custom loading script. The 'audio' feature needs to correctly map to audio files.")
    print(f"Example conceptual loading (might need a custom script or specific data_files pattern):")
    # dataset = load_dataset("json", data_files={"train": f"hf://datasets/{dataset_id}/metadata_entries/*.json"})
    # Then map audio files, e.g.:
    # def map_audio(batch):
    #     batch["audio"] = [f"hf://datasets/{dataset_id}/{path}" for path in batch["hf_audio_path"]]
    # return batch
    # dataset = dataset.map(map_audio, batched=True)
    # dataset = dataset.cast_column("audio", Audio(sampling_rate=16000)) # Example cast

except Exception as e:
    print(f"Error loading dataset: {e}")
    print("Please refer to the dataset repository for specific instructions on loading based on its file structure.")

# Example of accessing an item (assuming dataset is loaded)
# if 'dataset' in locals() and len(dataset['train']) > 0:
#     sample = dataset["train"][0]
#     print(sample["text"])
#     print(sample["audio"])