snorbyte/indic-text-audio-sample

Dataset Card for Indic Text Audio Sample Dataset

Dataset Details

Dataset Description

The IndicTextAudioSample Dataset is a multilingual, text-speech pair sample dataset. It features human-voiced recordings of dialogues in nine Indian languages: Hindi, Tamil, Telugu, Punjabi, Malayalam, Kannada, Bengali, Gujarati, and Marathi.

• Curated by: snorbyte
• Funded by: snorbyte
• Shared by: snorbyte
• Language(s) (NLP): hi, ta, te, pa, ml, kn, bn, gu, mr
• License: CC BY 4.0

Dataset Sources

• Repository: IndicTextAudioSample

Code

pip install huggingface_hub pandas pyarrow

import base64
import tempfile
import wave

from huggingface_hub import hf_hub_download
import pandas as pd

# Download the dataset file from Hugging Face
repo_id = "snorbyte/indic-text-audio-sample"
filename = "data_shard_000_zstd.parquet"
local_file = hf_hub_download(repo_id=repo_id, filename=filename, repo_type="dataset")
print("Downloaded to:", local_file)


# Load the Parquet file and get the first row
df = pd.read_parquet(local_file)
row = df.iloc[0]
print(row)

# Save the audio to a temporary WAV file
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as f:
    audio_bytes = row["audio"]["bytes"]  # extract raw bytes
    f.write(audio_bytes)
    temp_audio_path = f.name

print("Audio saved to:", temp_audio_path)

Uses

Direct Use

The dataset is intended for a wide range of applications, including but not limited to:

• Automatic Speech Recognition (ASR): Training and evaluating systems that transcribe spoken language into text.
• Text-to-Speech (TTS): Synthesizing natural-sounding speech from text prompts and evaluating TTS models.
• Multilingual Modeling: Developing models that generalize across languages for both speech and text processing.
• Demographic-Aware Modeling: Using age and gender metadata to develop or audit models for fairness, personalization, and bias analysis.
• Voice Cloning and Speech Synthesis: Training or evaluating models for voice conversion and synthesis using speaker-specific audio samples.
• Audio Classification: Classifying attributes such as speaker gender, age group from audio signals.
• Language Identification: Determining the spoken language from an audio sample.

Out-of-Scope Use

• Any use in sensitive applications like medical, legal, or surveillance without rigorous validation.
• Any use that attempts to infer personal attributes beyond what’s provided (age/gender).
• Generation or impersonation of real people using synthesized speech from dataset samples.

Dataset Structure

Each record in the dataset corresponds to a single text transcript and audio recording pair, along with user metadata. The dataset includes:

General Information

• language: Language used in the text and audio recording.
• audio: Complete conversation audio file in raw bytes.
• text: Text transcript of the audio.
• user_age: Age of the speaker.
• user_gender: Gender of the speaker.

Sample

language	audio	text	user_age	user_gender
hindi	bytes	जीत-हार के सपने में खोये प्रत्याशी	20.0	man

The sample dataset includes ~ 100 hours of audio and its equivalent text transcripts.

It comprises approximately 53.8% male and 46.2% female speakers, with 50% of the data contributed by individuals aged 18–30.

The following table shows the number of text-audio-recordings by language.

Language	Count
Hindi	5131
Tamil	5356
Gujarati	5576
Kannada	5308
Bengali	5752
Punjabi	5044
Telugu	5259
Marathi	5752
Malayalam	5148

Source Data

Purpose

This dataset aims to accelerate the development of language technologies in the Indic ecosystem by providing accessible and diverse resources.

Who are the source data producers?

All speakers voluntarily participated in the project and were compensated for their audio recordings. They represented a diverse range of age groups, genders, and professions.

Personal and Sensitive Information

• No personally identifiable information (PII) is present.
• Only age (grouped) and gender metadata are retained.
• All user IDs are anonymized.

Recommendations

• Supplement with additional datasets to improve dialect and age diversity.
• Validate model behavior across all demographic segments.
• Avoid over-interpreting demographic signals unless explicitly modeled and evaluated.

Citation

BibTeX:

@misc{indictextaudio2025,
  title={IndicTextAudio Sample Dataset},
  author={snorbyte},
  year={2025},
  howpublished={\url{https://huggingface.co/datasets/snorbyte/indic-text-audio-sample}},
  note={CC-BY 4.0}
}