app.py

import warnings
warnings.filterwarnings("ignore")

import os
import re
import pywt
import librosa
import webrtcvad
import nbimporter
import torchaudio
import numpy as np
import gradio as gr
import scipy.signal
import soundfile as sf
from scipy.io.wavfile import write
from transformers import pipeline
from transformers import AutoProcessor
from pyctcdecode import build_ctcdecoder
from transformers import Wav2Vec2ProcessorWithLM
from text2int import text_to_int
from isNumber import is_number
from Text2List import text_to_list
from convert2list import convert_to_list
# from processDoubles import process_doubles
from replaceWords import replace_words
from applyVad import apply_vad
from wienerFilter import wiener_filter
from highPassFilter import high_pass_filter
from waveletDenoise import wavelet_denoise
from scipy.signal import butter, lfilter, wiener

asr_model_malayalam = pipeline("automatic-speech-recognition", model="cdactvm/w2v-bert-malayalam")

def createlex(filename):
#filename = "num_map.txt"

# Initialize an empty dictionary
    data_dict = {}

# Open the file and read it line by line
    with open(filename, "r", encoding="utf-8") as f:
        for line in f:
        # Strip newline characters and split by tab
            key, value = line.strip().split("\t")
        # Add to dictionary
            data_dict[key] = value
    return data_dict

tellex=createlex("num_words_tel.txt")   
kanlex=createlex("num_words_kn.txt")
def addnum(inlist):
    sum=0
    for num in inlist:
        sum+=int(num)
   
    return sum

from rapidfuzz import process
def get_val(word, lexicon):
    threshold = 80  # Minimum similarity score
    length_difference = 4
    #length_range = (4, 6)  # Acceptable character length range (min, max)

    # Find the best match above the similarity threshold
    result = process.extractOne(word, lexicon.keys(), score_cutoff=threshold)
    print (result)
    if result:
        match, score, _ = result
        #print(lexicon[match])
        #return lexicon[match]
        if abs(len(match) - len(word)) <= length_difference:
        #if length_range[0] <= len(match) <= length_range[1]:
            return lexicon[match]
        else:
            return None
    else:
        return None
def convert2numtel(input, lex):
    input += " #"  # Add a period for termination
    words = input.split()
    i = 0
    num = 0
    outstr = ""
    digit_end = True
    numlist = []
    addflag = False
    prevword=""
    single_list=[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,15,17,18,19]
    # Process the words
    while i < len(words):
        #checkwordlist = handleSpecialnum(words[i])
        
        # Handle special numbers
        #if len(checkwordlist) == 2:
        #    words[i] = checkwordlist[0]
        #    words.insert(i + 1, checkwordlist[1])  # Collect new word for later processing

        # Get numerical value of the word
        numval = get_val(words[i], lex)
        if numval is not None:
            if prevword not in single_list:
                addflag = True
                numlist.append(numval)
            else:
                if addflag:
                    numlist.append(numval)
                    num = addnum(numlist)
                    outstr += str(num) + " "
                    addflag = False
                    numlist = []
                else:
                    outstr += " " + str(numval) + " "
            digit_end = False
            prevword=numval
        else:
            prevword=""
            if addflag:
                num = addnum(numlist)
                outstr += str(num) + " " + words[i] + " "
                addflag = False
                numlist = []
            else:
                outstr += words[i] + " "
            if not digit_end:
                digit_end = True
        
        # Move to the next word
        i += 1

    # Final processing
    outstr = outstr.replace('#','')  # Remove trailing spaces
    return outstr

def convert2numkn(input, lex):
    input += " ######"  # Add a period for termination
    words = input.split()
    i = 0
    num = 0
    outstr = ""
    digit_end = True
    numlist = []
    addflag = False
    
    prevword = []
    
    # Process the words
    while i < len(words):
        
        # Get numerical value of the word
        numval = get_val(words[i], lex)
        if len(prevword)>=3:
            prevword.pop(0)
            prevword.append(words[i])
        else:
            prevword.append(words[i])
        if numval is not None:
           
            addflag = True
           
            numlist.append(numval)
           
        else:
           
            
            #print("word--->"+words[i])
            #print("addflagword--->"+str(addflag))
            prevwords=" ".join(prevword)
            #print("prev word--->"+prevwords)
            numval=get_val(prevwords,lex)
            if numval is not None:
                #addflag=True
                #print("numval " +numval)
                numlist=[]
                #print("First outstr--->"+outstr)
                
                
                outwords = outstr.split()
                outstr=" ".join(outwords[:-1])
                #print("outstr--->"+outstr)
                
                outstr += " " + str(numval) + " " 
                #print(" aoutstr--->"+outstr)
                numval=0
                addflag=False
                
            else:
                if addflag:
                    num = addnum(numlist)
                    outstr += str(num) + " " + words[i] + " "
                    #print("penlast outstr--->"+outstr)
                    addflag = False
                    numlist = []
                else:
                    outstr += words[i] + " "
                    #print("last outstr--->"+outstr)
                if not digit_end:
                    digit_end = True
        
        
       
        # Move to the next word
        i += 1

    # Final processing
    outstr = outstr.replace('#','')  # Remove trailing spaces
    return outstr
    
# Function to apply a high-pass filter
def high_pass_filter(audio, sr, cutoff=300):
    nyquist = 0.5 * sr
    normal_cutoff = cutoff / nyquist
    b, a = butter(1, normal_cutoff, btype='high', analog=False)
    filtered_audio = lfilter(b, a, audio)
    return filtered_audio

# Function to apply wavelet denoising
def wavelet_denoise(audio, wavelet='db1', level=1):
    coeffs = pywt.wavedec(audio, wavelet, mode='per')
    sigma = np.median(np.abs(coeffs[-level])) / 0.5
    uthresh = sigma * np.sqrt(2 * np.log(len(audio)))
    coeffs[1:] = [pywt.threshold(i, value=uthresh, mode='soft') for i in coeffs[1:]]
    return pywt.waverec(coeffs, wavelet, mode='per')

# Function to apply a Wiener filter for noise reduction
def apply_wiener_filter(audio):
    return wiener(audio)

    
# Function to handle speech recognition
def recognize_speech_malayalam_model1(audio_file):
    audio, sr = librosa.load(audio_file, sr=16000)
    audio = high_pass_filter(audio, sr)
    audio = apply_wiener_filter(audio)
    denoised_audio = wavelet_denoise(audio)
    result = asr_model_malayalam(denoised_audio)
    text_value = result['text']
    converted_to_list = convert_to_list(cleaned_text, text_to_list())
    replaced_words = replace_words(converted_to_list)
    converted_text = text_to_int(replaced_words)
    return text_value +" -----------------> " + converted_text

#     # Function to handle speech recognition
# def recognize_speech_malayalam2(audio_file):
#     audio, sr = librosa.load(audio_file, sr=16000)
#     audio = high_pass_filter(audio, sr)
#     audio = apply_wiener_filter(audio)
#     denoised_audio = wavelet_denoise(audio)
#     result = asr_model_malayalam(denoised_audio)
#     text_value = result['text']
#     # cleaned_text = text_value.replace("[UNK]", "")
#     # converted_text=convert2numkn(cleaned_text,kanlex)
#     return cleaned_text +" -----------------> " + converted_text

def sel_lng(lng, mic=None, file=None):
    if mic is not None:
        audio = mic
    elif file is not None:
        audio = file
    else:
        return "You must either provide a mic recording or a file"
    
    if lng == "malayalam_model1":
        return recognize_speech_malayalam_model1(audio)
    # elif lng == "malayalam_model2":
    #     return recognize_speech_malayalam_model2(audio)

             
demo=gr.Interface(
    fn=sel_lng, 
      
    inputs=[
        gr.Dropdown([
            "malayalam_model1"],label="Select Model"),
        gr.Audio(sources=["microphone","upload"], type="filepath"),
    ],
    outputs=[
        "textbox"
    ],
    title="Automatic Speech Recognition",
    description = "Demo for Automatic Speech Recognition. Use microphone to record speech. Please press Record button. Initially it will take some time to load the model. The recognized text will appear in the output textbox",
      ).launch()