app.py

"""
Script to translate given single english audio file to corresponding hindi text
Usage : python s2t_en2hi.py <audio_file_path> <averaged_checkpoints_file_path>
"""



import gradio as gr
import sys
import os
import subprocess
from pydub import AudioSegment
from huggingface_hub import snapshot_download

def install_fairseq():
    try:
        # Run pip install command to install fairseq
        subprocess.check_call(["pip", "install", "fairseq"])
        subprocess.check_call(["pip", "install", "sentencepiece"])
        subprocess.check_call(["pip", "install", "soundfile"])
        subprocess.check_call(["pip", "install", "gTTS"])
        subprocess.check_call(["pip", "install", "gtts"])
        return "fairseq successfully installed!"
    except subprocess.CalledProcessError as e:
        return f"An error occurred while installing fairseq: {str(e)}"


from gtts import gTTS

def convert_audio_to_16k_wav(audio_input):
    sound = AudioSegment.from_file(audio_input)
    sample_rate = sound.frame_rate
    num_channels = sound.channels
    num_frames = int(sound.frame_count())
    filename = audio_input.split("/")[-1]
    print("original file is at:", audio_input)
    if (num_channels > 1) or (sample_rate != 16000): # convert to mono-channel 16k wav
        if num_channels > 1:
            sound = sound.set_channels(1)
        if sample_rate != 16000:
            sound = sound.set_frame_rate(16000)
        num_frames = int(sound.frame_count())
        filename = filename.replace(".wav", "") + "_16k.wav"
        sound.export(f"{filename}", format="wav")
    return filename


def run_my_code(input_text, language):
    # TODO better argument handling
    audio=convert_audio_to_16k_wav(input_text)
    hi_wav = audio

    data_root=""
    model_checkpoint=""
    d_r=""
    lang=''

    if(language=="Hindi"):
        model_checkpoint = "./models/hi_m.pt"
        data_root="./lang/hi/"
        lang='hi'

    if(language=="Gujrati"):
        model_checkpoint = "./models/gj_m.pt"
        data_root="./lang/gj/"
        lang='gu'

    if(language=="Bengali"):
        model_checkpoint = "./models/bn_m.pt"
        data_root="./lang/bn/"      
        lang='bn'  

    if(language=="Nepali"):
        model_checkpoint = "./models/ne_m.pt"
        data_root="./lang/ne/"
        lang='ne'

    if(language=="Tamil"):
        model_checkpoint = "./models/tm_m.pt"
        data_root="./lang/tm/"
        lang='ta'

    if(language=="Marathi"):
        model_checkpoint = "./models/mt_m.pt"
        data_root="./lang/mt/" 
        lang='mr'       
        
    #os.system(f"cp {hi_wav} {data_root}data/tst-COMMON/wav/test.wav")
    f = open('input.txt', 'w')
    f.write(hi_wav)

    f = open('input.txt', 'r')
    content = f. read()
    print(content)
    print(hi_wav)

    print("------Performing translation...")
    #subprocess.run(["fairseq-interactive", data_root, "--config-yaml", "config_st.yaml", "--task", "speech_to_text", "--path", model_checkpoint, "--max-tokens", "50000", "--beam", "5" ,"--input" ,"input.txt"])
    translation_result = subprocess.run(["fairseq-interactive", data_root, "--config-yaml", "config_st.yaml", "--task", "speech_to_text", "--path", model_checkpoint, "--max-tokens", "50000", "--beam", "5" ,"--input" ,"input.txt"], capture_output=True, text=True)
    translation_result_text = translation_result.stdout
 
    lines = translation_result_text.split("\n")

    output_text=""
    print("\n\n------Translation results are:")
    for i in lines:
        if (i.startswith("D-0")):
            print(i.split("\t")[2])
            output_text=i.split("\t")[2]
            break

    output_audio = "output_audio.mp3"
    tts = gTTS(text=output_text, lang=lang)
    tts.save(output_audio) 



    #os.system(f"rm test.wav")
    f = open('input.txt', 'w')
    f.write("")

    f = open('input.txt', 'r')
    content = f. read()
    print(content)
    return output_text, output_audio

install_fairseq()

# Define the input and output interfaces for Gradio
#inputs = [
  #      gr.inputs.Audio(source="microphone", type="filepath", label="Record something (in English)..."),
  #      gr.inputs.Dropdown(list(LANGUAGE_CODES.keys()), default="Hindi", label="From English to Languages X..."),
   # ]

#input_textbox = gr.inputs.Textbox(label="test2.wav")
#input=gr.inputs.Audio(source="microphone", type="filepath", label="Record something (in English)...")
#audio=convert_audio_to_16k_wav(input)
output_textbox = gr.outputs.Textbox(label="Translated Text")

# Create a Gradio interface
iface = gr.Interface(
        fn=run_my_code, 
        inputs=[gr.inputs.Audio(source="microphone", type="filepath", label="Record something (in American English accent)"), gr.inputs.Radio(["Hindi", "Gujrati", "Bengali", "Tamil", "Nepali", "Marathi"], label="Language")], 
        outputs=[output_textbox, gr.outputs.Audio(label="Output speech", type="filepath", live=True)],
        title="English to Indic Language Translator")

# Launch the interface
iface.launch()