src/pronunciation.py

import torch
import librosa
import numpy as np
import os
import traceback
import subprocess
import shutil

from transformers import (
    Wav2Vec2ForCTC, 
    AutoTokenizer, 
    Wav2Vec2FeatureExtractor
)

print("Loading Pronunciation module...")

MODEL_ID = "facebook/wav2vec2-lv-60-espeak-cv-ft"
model = None
tokenizer = None
feature_extractor = None

def find_espeak_exe():
    candidates = [
        r"C:\Program Files\eSpeak NG\espeak-ng.exe",
        r"C:\Program Files (x86)\eSpeak NG\espeak-ng.exe",
        r"D:\Program Files\eSpeak NG\espeak-ng.exe"
    ]
    path_in_env = shutil.which("espeak-ng")
    if path_in_env: return path_in_env
    
    for path in candidates:
        if os.path.exists(path):
            return path
    return None

ESPEAK_PATH = find_espeak_exe()
if ESPEAK_PATH:
    print(f"Found eSpeak at: {ESPEAK_PATH}")
else:
    print("WARNING: eSpeak-ng not found. IPA generation will fail.")

try:
    print("Loading Feature Extractor...")
    feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(MODEL_ID)
    
    print("Loading Tokenizer...")
    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
    
    print("Loading Acoustic Model...")
    model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)
    
    print("Pronunciation module ready.")
except Exception as e:
    print(f"Failed to load AI model: {e}")

def get_expected_ipa(text):
    """Gọi subprocess espeak-ng.exe để lấy IPA chuẩn từ văn bản."""
    if not ESPEAK_PATH:
        return "N/A"
    
    try:
        cmd = [ESPEAK_PATH, "-v", "en-us", "-q", "--ipa", text]
        
        startupinfo = None
        if os.name == 'nt':
            startupinfo = subprocess.STARTUPINFO()
            startupinfo.dwFlags |= subprocess.STARTF_USESHOWWINDOW
            
        result = subprocess.run(
            cmd, 
            capture_output=True, 
            text=True, 
            encoding='utf-8',
            startupinfo=startupinfo
        )
        
        if result.returncode == 0:
            return result.stdout.strip().replace('\n', ' ')
        else:
            return "N/A"
            
    except Exception as e:
        print(f"Subprocess error: {e}")
        return "N/A"

def grade_pronunciation_advanced(audio_path, reference_text):
    """
    Trả về chuỗi IPA thực tế (Audio) và IPA chuẩn (Text).
    """
    actual_ipa = "N/A"
    if model and tokenizer and feature_extractor:
        try:
            y, sr = librosa.load(audio_path, sr=16000)
            
            input_values = feature_extractor(y, sampling_rate=16000, return_tensors="pt").input_values
            with torch.no_grad():
                logits = model(input_values).logits
            
            predicted_ids = torch.argmax(logits, dim=-1)
            actual_ipa = tokenizer.batch_decode(predicted_ids, skip_special_tokens=True)[0]
            
        except Exception as e:
            print(f"AI IPA Error: {e}")
            actual_ipa = "Error"

    expected_ipa = get_expected_ipa(reference_text)

    return {
        "actual_ipa": actual_ipa,
        "expected_ipa": expected_ipa
    }