logic.py

from phonemizer.separator import Separator
from phonemizer import phonemize, backend
# from phonemizer.backend.espeak.wrapper import EspeakWrapper
from Levenshtein import distance as levenshtein_distance

import whisper 
import torch 
if not backend.EspeakBackend.is_available():
    import os
    os.system('sudo apt-get install espeak-ng')
device = 'cuda:0' if torch.cuda.is_available() else 'cpu'

model = whisper.load_model("base.en", device=device)
separator = Separator(phone=None, word=' | ',)

# EspeakWrapper.set_library(r"C:\Program Files\eSpeak NG\libespeak-ng.dll")

def transcribe(audio):
    result = model.transcribe(audio, word_timestamps=False, no_speech_threshold=0.4,  compression_ratio_threshold=2, temperature=0)
    return {'language': result['language'], 'text': result['text']}

def text2phoneme(text):
    return phonemize(text.lower(), backend='espeak' , separator=separator, strip=True, with_stress=False, tie=False, language='en-us')

def rate_pronunciation(expected_phonemes, actual_phonemes):
    expected_phonemes = expected_phonemes.split(" | ")
    actual_phonemes = actual_phonemes.split(" | ")
    # Calculate the Levenshtein distance between the two phoneme sequences
    results = []
    for i, base_word in enumerate(actual_phonemes):
        best_dist = float('inf')
        error_threshold = len(base_word) * 0.45
        for pred_word_id in range(max(0, i-2), i + min(6, len(expected_phonemes) - i)):
            dist = levenshtein_distance(expected_phonemes[pred_word_id], base_word,)
            if dist < best_dist:
                best_dist = dist
            if best_dist == 0:  # Early stopping on perfect match
                break
        if best_dist == 0:
           results.append(3) 
        elif best_dist <= error_threshold:
            results.append(2) 
        else:
            results.append(1) 
    return results

def compare_audio_with_text(audio, text):
    transcribtion = transcribe(audio)['text']
    print(transcribtion)
    transcribtion = text2phoneme(transcribtion)
    text_phone    = text2phoneme(text)
    scores        = rate_pronunciation(transcribtion, text_phone)

    result = [(word, s) for word, s in zip(text.split(), scores)]
    return result

if __name__ == '__main__':
    
    text = 'i have ADHD '
    text = text2phoneme(text)
    file_path = r'user_recording.wav'
    trans = transcribe(file_path)['text']
    print(trans)
    trans = text2phoneme(trans)
    print('base:', text)
    print('predicted:', trans)
    result = rate_pronunciation(trans, text)
    print(result)