app_api.py

from fastapi import FastAPI, File, UploadFile, Form
from fastapi.responses import JSONResponse
import uvicorn
import tempfile
import nemo.collections.asr as nemo_asr
import re
import os
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch
from word2number import w2n
from deep_translator import GoogleTranslator

# ===== Arabic number mapping (expanded) =====
arabic_numbers = {
    # Basic digits
    "صفر": "0", "زيرو": "0", "٠": "0","زيو": "0","زير": "0","زر": "0","زروا": "0","زرا": "0","زيره ": "0","زرو ": "0",
    "واحد": "1", "واحدة": "1", "١": "1",
    "اتنين": "2", "اثنين": "2", "إثنين": "2", "اثنان": "2", "إثنان": "2", "٢": "2",
    "تلاتة": "3", "ثلاثة": "3", "٣": "3",
    "اربعة": "4", "أربعة": "4", "٤": "4",
    "خمسة": "5", "٥": "5",
    "ستة": "6", "٦": "6",
    "سبعة": "7", "٧": "7",
    "تمانية": "8", "ثمانية": "8", "٨": "8",
    "تسعة": "9", "٩": "9",

    # Teens
    "عشرة": "10", "١٠": "10",
    # 11
    "احد عشر": "11", "واحد عشر": "11", "حداشر": "11",
    "١ عشر": "11", "1 عشر": "11", "١عشر": "11", "1عشر": "11",
    "١١": "11", "11": "11",

    # 12
    "اثنا عشر": "12", "اثني عشر": "12", "اتناشر": "12",
    "٢ عشر": "12", "2 عشر": "12", "٢عشر": "12", "2عشر": "12",
    "١٢": "12", "12": "12",

    # 13
    "ثلاثة عشر": "13", "تلاتة عشر": "13", "تلتاشر": "13",
    "٣ عشر": "13", "3 عشر": "13", "٣عشر": "13", "3عشر": "13",
    "١٣": "13", "13": "13",

    # 14
    "أربعة عشر": "14", "اربعة عشر": "14", "اربعتاشر": "14",
    "٤ عشر": "14", "4 عشر": "14", "٤عشر": "14", "4عشر": "14",
    "١٤": "14", "14": "14",

    # 15
    "خمسة عشر": "15", "خمسه عشر": "15", "خمستاشر": "15",
    "٥ عشر": "15", "5 عشر": "15", "٥عشر": "15", "5عشر": "15",
    "١٥": "15", "15": "15",

    # 16
    "ستة عشر": "16", "سته عشر": "16", "ستاشر": "16",
    "٦ عشر": "16", "6 عشر": "16", "٦عشر": "16", "6عشر": "16",
    "١٦": "16", "16": "16",

    # 17
    "سبعة عشر": "17", "سبعه عشر": "17", "سبعتاشر": "17",
    "٧ عشر": "17", "7 عشر": "17", "٧عشر": "17", "7عشر": "17",
    "١٧": "17", "17": "17",

    # 18
    "ثمانية عشر": "18", "تمانية عشر": "18", "طمنتاشر": "18",
    "٨ عشر": "18", "8 عشر": "18", "٨عشر": "18", "8عشر": "18",
    "١٨": "18", "18": "18",

    # 19
    "تسعة عشر": "19", "تسعه عشر": "19", "تسعتاشر": "19",
    "٩ عشر": "19", "9 عشر": "19", "٩عشر": "19", "9عشر": "19",
    "١٩": "19", "19": "19",

    # Tens
    "عشرين": "20", "٢٠": "20",
    "تلاتين": "30", "ثلاثين": "30", "٣٠": "30",
    "اربعين": "40", "أربعين": "40", "٤٠": "40",
    "خمسين": "50", "٥٠": "50",
    "ستين": "60", "٦٠": "60",
    "سبعين": "70", "٧٠": "70",
    "تمانين": "80", "ثمانين": "80", "٨٠": "80","تمانون": "80","ثمانون": "80",
    "تسعين": "90", "٩٠": "90",

    # Hundreds
    "مية": "100", "مائة": "100", "مئة": "100", "١٠٠": "100",
    "ميتين": "200", "مائتين": "200",
    "تلاتمية": "300", "ثلاثمائة": "300",
    "اربعمية": "400", "أربعمائة": "400",
    "خمسمية": "500", "خمسمائة": "500",
    "ستمية": "600", "ستمائة": "600",
    "سبعمية": "700", "سبعمائة": "700",
    "تمانمية": "800", "ثمانمائة": "800",
    "تسعمية": "900", "تسعمائة": "900",

    # Thousands
    "ألف": "1000", "الف": "1000", "١٠٠٠": "1000",
    "ألفين": "2000", "الفين": "2000",
    "تلات تلاف": "3000", "ثلاثة آلاف": "3000",
    "اربعة آلاف": "4000", "أربعة آلاف": "4000",
    "خمسة آلاف": "5000",
    "ستة آلاف": "6000",
    "سبعة آلاف": "7000",
    "تمانية آلاف": "8000", "ثمانية آلاف": "8000",
    "تسعة آلاف": "9000",

    # Large numbers
    "عشرة آلاف": "10000",
    "مية ألف": "100000", "مائة ألف": "100000",
    "مليون": "1000000", "١٠٠٠٠٠٠": "1000000",
    "ملايين": "1000000",
    "مليار": "1000000000", "١٠٠٠٠٠٠٠٠٠": "1000000000",
    # ===== Compound tens (Arabic + digit forms) =====
    "واحد وعشرون": "21", "1 وعشرون": "21",
    "اثنان وعشرون": "22", "٢ وعشرون": "22",
    "ثلاثة وعشرون": "23", "٣ وعشرون": "23",
    "اربعة وعشرون": "24", "٤ وعشرون": "24",
    "خمسة وعشرون": "25", "٥ وعشرون": "25",
    "ستة وعشرون": "26", "٦ وعشرون": "26",
    "سبعة وعشرون": "27", "٧ وعشرون": "27",
    "تمانية وعشرون": "28", "ثمانية وعشرون": "28", "٨ وعشرون": "28",
    "تسعة وعشرون": "29", "٩ وعشرون": "29",

    "ثمانية وثمانون": "88", "8 وثمانون": "88",
    "اثنان وثمانون": "82", "٢ وثمانون": "82",
    "خمسة وستون": "65", "5 وستون": "65",
    "ستة عشر": "16", "٦ عشر": "16",
    "اثنا عشر": "12", "١٢": "12",
    "ثلاثة وثلاثون": "33", "٣٣": "33", "33": "33",
    "أربعة وأربعون": "44", "٤٤": "44", "44": "44",
    "خمسة وخمسون": "55", "٥٥": "55", "55": "55",
    "ستة وستون": "66", "٦٦": "66", "66": "66",
    "سبعة وسبعون": "77", "٧٧": "77", "77": "77",
    "ثمانية وثمانون": "88", "٨٨": "88", "88": "88",
    "تسعة وتسعون": "99", "٩٩": "99", "99": "99",
}

def replace_arabic_numbers(text: str) -> str:
    for word, digit in arabic_numbers.items():
        text = re.sub(rf"\b{word}\b", digit, text)
    return text


# ===== FastAPI app =====
app = FastAPI(title="Arabic ASR API", description="ASR API with NeMo and Arabic digit conversion")

# Load model once on startup
@app.on_event("startup")
def load_model():
    global asr_model
    global model
    global tokenizer
    global device
    #model_path = os.getenv("NEMO_MODEL_PATH", "C:/Users/thegh/Python_Projects/Expertflow/UnderProgress/Peter_Projects/nvidia_asr_eg_conformer_better_than_whisper/stt_ar_fastconformer_hybrid_large_pcd_v1.0.nemo") 
    model_path = "C:/Users/thegh/Python_Projects/Expertflow/UnderProgress/Peter_Projects/NP_Detection_Nvidia_conformer/stt_ar_fastconformer_hybrid_large_pc_v1.0.nemo"
    asr_model = nemo_asr.models.EncDecCTCModel.restore_from(model_path)
    # Load once globally
    # model_name = "alaasayed_ai/Egyptian_Arabic_to_English"
    model_translator_name = "ukaAi/Egyptian_dialect_to_arabic"
    tokenizer = AutoTokenizer.from_pretrained(model_translator_name)
    model = AutoModelForSeq2SeqLM.from_pretrained(model_translator_name)

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = model.to(device)
def translate_egyptian_to_english(text: str) -> str:
    """
    Translates Egyptian Arabic text to English using the fine-tuned NLLB model.
    
    Parameters:
    - text (str): The input Egyptian Arabic text
    
    Returns:
    - str: The translated English text
    """
    tokenizer.src_lang = "arz_Arab"
    forced_bos_token_id = tokenizer.convert_tokens_to_ids("eng_Latn")

    inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True, max_length=512)
    inputs = {k: v.to(device) for k, v in inputs.items()}

    translated = model.generate(
        **inputs,
        forced_bos_token_id=forced_bos_token_id,
        max_length=512,
        num_beams=4,
        early_stopping=True
    )

    return tokenizer.decode(translated[0], skip_special_tokens=True)
# Cardinal and tens
WORD_TO_NUM = {
    "zero": 0, "one": 1, "two": 2, "three": 3, "four": 4, "five": 5,
    "six": 6, "seven": 7, "eight": 8, "nine": 9, "ten": 10,
    "eleven": 11, "twelve": 12, "thirteen": 13, "fourteen": 14,
    "fifteen": 15, "sixteen": 16, "seventeen": 17, "eighteen": 18,
    "nineteen": 19, "twenty": 20, "thirty": 30, "forty": 40,
    "fifty": 50, "sixty": 60, "seventy": 70, "eighty": 80, "ninety": 90
}

# Ordinals
ORDINAL_TO_NUM = {
    "first": 1, "second": 2, "third": 3, "fourth": 4, "fifth": 5,
    "sixth": 6, "seventh": 7, "eighth": 8, "ninth": 9, "tenth": 10,
    "eleventh": 11, "twelfth": 12, "thirteenth": 13, "fourteenth": 14,
    "fifteenth": 15, "sixteenth": 16, "seventeenth": 17, "eighteenth": 18,
    "nineteenth": 19, "twentieth": 20, "thirtieth": 30, "fortieth": 40,
    "fiftieth": 50, "sixtieth": 60, "seventieth": 70, "eightieth": 80, "ninetieth": 90
}

def normalize_token(token: str):
    """Convert a single token or hyphenated token into a number if possible."""
    token = token.lower()

    # Handle ordinals
    if token in ORDINAL_TO_NUM:
        return ORDINAL_TO_NUM[token]

    # Handle hyphenated compounds like 'thirty-nine'
    if "-" in token:
        parts = token.split("-")
        nums = [WORD_TO_NUM.get(p) for p in parts if p in WORD_TO_NUM]
        if nums:
            return sum(nums)

    # Handle normal cardinals
    return WORD_TO_NUM.get(token)


def words_to_numbers(phrase: str):
    tokens = phrase.lower().strip().split()
    nums = [normalize_token(t) for t in tokens if normalize_token(t) is not None]

    if not nums:
        return []

    # Case: three tokens like "two one ninety" → 91
    if len(nums) == 3:
        return [int(f"{nums[0]}{nums[1]}") + nums[2]]

    # Case: two tokens like "five thirty" → 35
    if len(nums) == 2:
        if nums[1] >= 20:
            return [nums[0] + nums[1]]
        else:
            return [int("".join(str(n) for n in nums))]

    # Otherwise, return each token separately
    return nums


def parse_numbers(text: str):
    chunks = re.split(r"[,\.;]", text)
    result = []
    for chunk in chunks:
        result.extend(words_to_numbers(chunk))
    return " ".join(str(n) for n in result)
@app.post("/transcribe")
async def transcribe_audio(file: UploadFile = File(...)):
    # Save uploaded file to a temp path
    with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp:
        tmp.write(await file.read())
        tmp_path = tmp.name

    try:
        # Run transcription
        result = asr_model.transcribe([tmp_path])
        print(result)
        raw_text = result[0].text
        print(raw_text)
       
        result = translate_egyptian_to_english(raw_text)

        print("\n=== English Translation ===\n")
        print(result)
        
        print(parse_numbers(result))
        # print (w2n.word_to_num(result))
        # Convert Arabic numbers
        # cleaned_text = replace_arabic_numbers(raw_text)
        # print("\n\n")
        # print(cleaned_text)
        # print("\n\n")
        return JSONResponse(content={"transcription": raw_text})

    finally:
        os.remove(tmp_path)


@app.post("/transcribe-bytes")
async def transcribe_audio_bytes(audio_bytes: bytes = File(...)):
    with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp:
        tmp.write(audio_bytes)
        tmp_path = tmp.name

    try:
        result = asr_model.transcribe([tmp_path])
        raw_text = result[0].text
        cleaned_text = replace_arabic_numbers(raw_text)

        return JSONResponse(content={"transcription": cleaned_text})

    finally:
        os.remove(tmp_path)


if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=8000, reload=True)