AJosh/audio-dataset

Audio Transcription Dataset

This dataset contains 197 audio recordings with their corresponding transcriptions for automatic speech recognition (ASR) tasks.

Dataset Description

This dataset includes:

• Audio files: High-quality voice recordings (.wav format)
• Transcriptions: Accurate text transcriptions of the spoken content
• Proper Audio feature type: Ready for model training (not just file paths!)

Dataset Statistics

• Total samples: 197
• Audio format: WAV files at 16kHz sampling rate
• Average transcript length: 56.6 characters
• Language: English

Sample Data

Audio File	Transcript
R1.wav	Hello, my name is Rocky.
R10.wav	I am speaking English for a voice recording.
R11.wav	This is a test sentence for training the model.

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("Aashish17405/audio-dataset")

# Access audio data (proper Audio type, not string!)
audio_sample = dataset['train'][0]['audio']
print(f"Sampling rate: {audio_sample['sampling_rate']}")
print(f"Audio array shape: {audio_sample['array'].shape}")
print(f"Transcript: {dataset['train'][0]['transcript']}")

# Ready for model training with transformers
from transformers import WhisperProcessor, WhisperForConditionalGeneration

processor = WhisperProcessor.from_pretrained("openai/whisper-tiny")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny")

# Process audio
inputs = processor(audio_sample["array"], sampling_rate=audio_sample["sampling_rate"], return_tensors="pt")

Features

✅ Proper Audio Type: Audio column shows as "Audio" feature, not "string"
✅ High Quality: Clear voice recordings
✅ Diverse Content: Various sentences and topics
✅ Training Ready: Formatted for immediate use with speech models

Use Cases

• Fine-tuning speech recognition models (Whisper, Wav2Vec2, etc.)
• Voice training and accent recognition
• Speech-to-text model development
• Audio processing research

License

MIT License - Free to use for research and commercial purposes.