Indian English Speech Emotion Dataset
Description
Overview Despite the global diversity of English accents, there is a notable scarcity of publicly available datasets for sentiment analysis focusing on Indian English speech. Existing speech emotion recognition datasets predominantly feature Western accents, limiting the development of region-specific models for Indian English, which is characterized by unique prosodic and phonetic variations. This dataset addresses this gap by providing a robust collection of Indian English speech samples tailored for emotion recognition, enabling advancements in culturally relevant AI applications. Research Hypothesis Acoustic features such as pitch, energy, and prosodic cues extracted directly from Indian English speech signals can effectively predict a speaker’s emotional state without requiring speech-to-text transcription. These signals contain distinguishable patterns for emotions like Happy, Sad, Angry, and Neutral, which can be learned by deep learning models such as Long Short-Term Memory (LSTM) networks. Data Collection and Gathering The dataset consists of audio recordings from native Indian English speakers across diverse regions of India, capturing a variety of accents and including both genders. Recordings were conducted in controlled, quiet environments to ensure high audio quality, with participants speaking short, natural sentences designed to express four emotions: Happy, Sad, Angry, and Neutral. The original dataset comprises 1,000 samples, evenly distributed across the four emotions . To enhance model robustness and generalizability, data augmentation techniques—including pitch shifting, time stretching, and controlled noise addition—were applied, expanding the dataset to 3,000 samples. The augmented dataset maintains balance across emotions and is organized into labeled folders (Happy, Sad, Angry, Neutral) to facilitate supervised learning. Notable Findings An LSTM-based model trained solely on acoustic features, without reliance on textual transcription, achieved a classification accuracy of 85% on the original dataset. With the inclusion of augmented data, the accuracy improved significantly to 96%. These results highlight the effectiveness of acoustic cues in capturing emotional states in Indian English speech and the important role of data augmentation in improving model performance. Data Interpretation and Usage This dataset is a valuable resource for researchers and practitioners in speech emotion recognition and related fields. Potential applications include: Developing and benchmarking transcription-free speech emotion recognition models. Analyzing acoustic and prosodic patterns unique to Indian English emotional speech. Building sentiment-aware applications such as voice assistants, call center analytics, and mental health monitoring tools. Investigating the impact of data augmentation on acoustic-based sentiment model performance.
Files
Steps to reproduce
Data Collection: Gather audio recordings from native Indian English speakers across India, ensuring diverse accents and genders. Record short sentences designed to express Happy, Sad, Angry, and Neutral emotions in a quiet environment using a microphone (e.g., USB or smartphone) in WAV format, yielding 1,000 samples (250 per emotion). Data Preprocessing: Normalize audio files and reduce noise using tools like noisereduce. Trim silence with librosa. Extract acoustic features (e.g., MFCCs) using librosa and standardize their dimensions. Organize files into folders labeled “Happy,” “Sad,” “Angry,” and “Neutral” with metadata in a CSV file. Data Augmentation: Apply techniques like pitch shifting, time stretching, and volume adjustment using librosa to expand the dataset to 3,000 samples (750 per emotion), maintaining balanced classes. Save augmented files in a separate directory with updated metadata. Model Training: Train an LSTM-based model on acoustic features using TensorFlow, with a train-validation-test split (e.g., 80-10-10). Use standard hyperparameters (e.g., Adam optimizer, categorical cross-entropy loss) and train for multiple epochs. Achieve 85% accuracy on the original 1,000 samples and 96% on the augmented 3,000 samples. Tools and Environment: Use Python, librosa, noisereduce, tensorflow, and pandas for processing and training. Record in a quiet setting with a standard microphone. Train on a laptop with GPU support.