Grammar-Scoring-Engine-for-Spoken-Data-Samples

This repository contains a pipeline to evaluate the grammatical accuracy of spoken audio using a fusion of ASR (Automatic Speech Recognition) and deep learning models. The solution transcribes audio using Wav2Vec2, preprocesses the text, and predicts grammar scores using a BERT + BiLSTM regression model.

---

Architecture

🚀 Overview

🧠 Task

Given an audio clip of a candidate speaking, predict a grammar score (1–5) reflecting their spoken grammar quality.

📊 Dataset

• train.csv: Contains audio file names + grammar labels.
• audios_train/: Folder of training .wav files.
• test.csv / audios_test/: Evaluation set without labels.

---

🛠️ Pipeline

1. 🎙️ Audio Transcription

• Model: facebook/wav2vec2-base-960h
• Output: Lowercased text using batch_decode.

2. 🧹 Text Preprocessing

• Remove punctuation, normalize spaces, lowercase.

3. 🧠 Modeling

• Architecture: BERT Embeddings → BiLSTM → Dense Regression Head
• Loss: MSELoss, optional Pearson/CCC loss
• Framework: PyTorch

4. 🧪 Evaluation

• Metric: Pearson Correlation on validation set.
• Visualization: Score distributions, loss curves.

---

📈 Results

Metric	Value
Pearson (val)	0.599
MSE (val)	5.78 50 epochs

---

🧪 Running the Pipeline

# Step 1: Texttranscription
python texttranscription.py

# Step 2: Preprocess text
python preprocess.py

# Step 3: Train model
python grammar_score_model.py

# Step 4: Run inference on test set
python inference.py