app.py

#!/usr/bin/env python
# coding: utf-8

# In[ ]:


# conver mp3 to wav 
# ffmpeg -i test_5.mp3 -b:a 16000 test_5.wav


# In[1]:


#Importing all the necessary packages
import nltk
import librosa
import IPython.display
import torch
import gradio as gr
from transformers import Wav2Vec2Tokenizer, Wav2Vec2ForCTC
nltk.download("punkt")


# In[2]:


#Loading the model and the tokenizer
model_name = "facebook/wav2vec2-base-960h"

#model_name = "facebook/wav2vec2-large-xlsr-53"
tokenizer = Wav2Vec2Tokenizer.from_pretrained(model_name)#omdel_name
model = Wav2Vec2ForCTC.from_pretrained(model_name)


# In[3]:


def load_data(input_file):
    """ Function for resampling to ensure that the speech input is sampled at 16KHz.
    """
    #read the file
    speech, sample_rate = librosa.load(input_file)
    #make it 1-D
    if len(speech.shape) > 1: 
        speech = speech[:,0] + speech[:,1]
    #Resampling at 16KHz since wav2vec2-base-960h is pretrained and fine-tuned on speech audio sampled at 16 KHz.
    if sample_rate !=16000:
        speech = librosa.resample(speech, sample_rate,16000)
    #speeches = librosa.effects.split(speech)
    return speech


# In[4]:


def correct_casing(input_sentence):
    """ This function is for correcting the casing of the generated transcribed text
    """
    sentences = nltk.sent_tokenize(input_sentence)
    return (' '.join([s.replace(s[0],s[0].capitalize(),1) for s in sentences]))


# In[5]:


def asr_transcript(input_file):
    """This function generates transcripts for the provided audio input
    """
    speech = load_data(input_file)
    #Tokenize
    input_values = tokenizer(speech, return_tensors="pt").input_values
    #Take logits
    logits = model(input_values).logits
    #Take argmax
    predicted_ids = torch.argmax(logits, dim=-1)
    #Get the words from predicted word ids
    transcription = tokenizer.decode(predicted_ids[0])
    #Output is all upper case
    transcription = correct_casing(transcription.lower())
    return transcription


# In[6]:


def asr_transcript_long(input_file,tokenizer=tokenizer, model=model ):
    transcript = ""
    # Ensure that the sample rate is 16k
    sample_rate = librosa.get_samplerate(input_file)

    # Stream over 30 seconds chunks rather than load the full file
    stream = librosa.stream(
        input_file,
        block_length=30,
        frame_length=sample_rate, #16000,
        hop_length=sample_rate, #16000
    )

    for speech in stream:
        if len(speech.shape) > 1:
            speech = speech[:, 0] + speech[:, 1]
        if sample_rate !=16000:
            speech = librosa.resample(speech, sample_rate,16000)
        input_values = tokenizer(speech, return_tensors="pt").input_values
        logits = model(input_values).logits

        predicted_ids = torch.argmax(logits, dim=-1)
        transcription = tokenizer.decode(predicted_ids[0])
        #transcript += correct_sentence(transcription.lower())
        transcript += correct_casing(transcription.lower())
        transcript += " "

    return transcript

from pydub import AudioSegment
from pydub.silence import split_on_silence
from pydub.playback import play 

sound = AudioSegment.from_file("./test_2.wav", format="wav")
chunks = split_on_silence(
    sound,

    # split on silences longer than 1000ms (1 sec)
    min_silence_len=5000,

    # anything under -16 dBFS is considered silence
    silence_thresh=-32, 

    # keep 200 ms of leading/trailing silence
    keep_silence=500
)#read the file
speech, sample_rate = librosa.load('./test_2.wav')
#make it 1-D
if len(speech.shape) > 1: 
    speech = speech[:,0] + speech[:,1]
#Resampling at 16KHz since wav2vec2-base-960h is pretrained and fine-tuned on speech audio sampled at 16 KHz.
if sample_rate !=16000:
    speech = librosa.resample(speech, sample_rate,16000)
part_of_speech = librosa.effects.split(speech)idx = -1
IPython.display.Audio(data=speech[part_of_speech[idx,0]:part_of_speech[idx,1]], rate=16000)
# In[ ]:


gr.Interface(asr_transcript_long,
             #inputs = gr.inputs.Audio(source="microphone", type="filepath", optional=True, label="Please record your voice"),
             inputs = gr.inputs.Audio(source="upload", type="filepath", optional=True, label="Upload your file here"),
             outputs = gr.outputs.Textbox(label="Output Text"),
             title="ASR using Wav2Vec 2.0",
             description = "This application displays transcribed text for given audio input",
             examples = [["Test_File1.wav"], ["Test_File2.wav"], ["Test_File3.wav"]], theme="grass").launch()


# In[ ]:





# In[ ]:





# In[ ]:





# In[7]:


#temp = asr_transcript_long('./test_2.wav')


# In[ ]:





# In[ ]: