HusseinBashir/Somali_TTS_API

1from fastapi import FastAPI, Request2from fastapi.responses import FileResponse3import torch4import numpy as np5import scipy.io.wavfile6from transformers import VitsModel, AutoTokenizer7import re8 9app = FastAPI()10 11# Load model and tokenizer12model = VitsModel.from_pretrained("Somali-tts/somali_tts_model")13tokenizer = AutoTokenizer.from_pretrained("saleolow/somali-mms-tts")14device = torch.device("cuda" if torch.cuda.is_available() else "cpu")15model.to(device)16model.eval()17 18number_words = {19    0: "eber", 1: "koow", 2: "labo", 3: "seddex", 4: "afar", 5: "shan",20    6: "lix", 7: "todobo", 8: "sideed", 9: "sagaal", 10: "toban",21    11: "toban iyo koow", 12: "toban iyo labo", 13: "toban iyo seddex",22    14: "toban iyo afar", 15: "toban iyo shan", 16: "toban iyo lix",23    17: "toban iyo todobo", 18: "toban iyo sideed", 19: "toban iyo sagaal",24    20: "labaatan", 30: "sodon", 40: "afartan", 50: "konton",25    60: "lixdan", 70: "todobaatan", 80: "sideetan", 90: "sagaashan",26    100: "boqol", 1000: "kun"27}28 29def number_to_words(number):30    number = int(number)31    if number < 20:32        return number_words[number]33    elif number < 100:34        tens, unit = divmod(number, 10)35        return number_words[tens * 10] + (" iyo " + number_words[unit] if unit else "")36    elif number < 1000:37        hundreds, remainder = divmod(number, 100)38        part = (number_words[hundreds] + " boqol") if hundreds > 1 else "boqol"39        if remainder:40            part += " iyo " + number_to_words(remainder)41        return part42    elif number < 1000000:43        thousands, remainder = divmod(number, 1000)44        words = [number_to_words(thousands) + " kun" if thousands != 1 else "kun"]45        if remainder:46            words.append("iyo " + number_to_words(remainder))47        return " ".join(words)48    else:49        return str(number)50 51def normalize_text(text):52    numbers = re.findall(r'\d+', text)53    for num in numbers:54        text = text.replace(num, number_to_words(num))55    text = text.replace("KH", "qa").replace("Z", "S")56    text = text.replace("SH", "SHa'a").replace("DH", "Dha'a")57    text = text.replace("ZamZam", "SamSam")58    return text59 60@app.post("/tts")61async def tts(request: Request):62    data = await request.json()63    text = normalize_text(data["text"])64    inputs = tokenizer(text, return_tensors="pt").to(device)65    with torch.no_grad():66        waveform = model(**inputs).waveform.squeeze().cpu().numpy()67    filename = "output.wav"68    scipy.io.wavfile.write(filename, rate=model.config.sampling_rate, data=(waveform * 32767).astype(np.int16))69    return FileResponse(filename, media_type="audio/wav")70