ReadAI - Pronunciation & Audio Quality Assessment Models

This repository contains two models for audio assessment:

1. Pronunciation Assessment Model (pronunciation_v3/)

A fine-tuned WhisperForAudioClassification model (based on openai/whisper-base) for binary pronunciation quality classification.

Labels

Label	ID
Bad	0
Good	1

Usage

from transformers import pipeline

classifier = pipeline(
    task="audio-classification",
    model="jecallora/readai",
    subfolder="pronunciation_v3"
)

result = classifier("audio_sample.wav")
print(result)
# [{'label': 'Good', 'score': 0.95}, {'label': 'Bad', 'score': 0.05}]

Model Details

• Architecture: WhisperForAudioClassification
• Base Model: openai/whisper-base
• Sampling Rate: 16,000 Hz
• Input Format: Audio (WAV, MP3, etc.)
• Framework: PyTorch (safetensors)

---

2. Audio Quality Classifier (audio_quality/)

A scikit-learn classifier for audio quality assessment.

Labels

Quality	Score
Very Good	100
Good	75
Bad	50
Very Bad	25

Files

• audio_classifier.joblib — Trained classifier
• scaler.joblib — StandardScaler for feature normalization
• label_encoder.joblib — Label encoder

Usage

import joblib
import librosa
import numpy as np

# Load models
classifier = joblib.load("audio_quality/audio_classifier.joblib")
scaler = joblib.load("audio_quality/scaler.joblib")
label_encoder = joblib.load("audio_quality/label_encoder.joblib")

# Extract features from audio (16kHz mono)
y, sr = librosa.load("audio_sample.wav", sr=16000, mono=True)

# Your feature extraction pipeline here...
# features = extract_features(y)
# scaled = scaler.transform([features])
# prediction = classifier.predict(scaled)
# label = label_encoder.inverse_transform(prediction)

Dependencies

• scikit-learn==1.5.0
• librosa==0.10.2.post1
• numpy==1.26.4
• joblib

---

Requirements

transformers>=4.41.2
torch>=2.3.1
torchaudio>=2.3.1
scikit-learn>=1.5.0
librosa>=0.10.2.post1
soundfile>=0.12.1
numpy>=1.26.4