app.py

import os

# Download and build ggergavos/whisper.cpp Kudos to this man for wonderful whisper implementation! 
# This means speed!

os.system('git clone https://github.com/ggerganov/whisper.cpp.git')
os.system('make -C ./whisper.cpp')

# Download models, add finetuned languages later once whisper finetuning event is ready
# Models are downloaded on the fly so we can get quite many models :) 

os.system('bash ./whisper.cpp/models/download-ggml-model.sh small')
os.system('bash ./whisper.cpp/models/download-ggml-model.sh base')
os.system('bash ./whisper.cpp/models/download-ggml-model.sh medium')
os.system('bash ./whisper.cpp/models/download-ggml-model.sh base.en')

#os.system('./whisper.cpp/main -m whisper.cpp/models/ggml-base.en.bin -f whisper.cpp/samples/jfk.wav')
#print("SEURAAVAKSI SMALL TESTI")
#os.system('./whisper.cpp/main -m whisper.cpp/models/ggml-small.bin -f whisper.cpp/samples/jfk.wav')
#print("MOI")



import gradio as gr
from pathlib import Path
import pysrt
import pandas as pd
import re
import time
import os
import json

from pytube import YouTube
from transformers import MarianMTModel, MarianTokenizer

import psutil
num_cores = psutil.cpu_count()
os.environ["OMP_NUM_THREADS"] = f"{num_cores}"
headers = {'Authorization': os.environ['DeepL_API_KEY']}

whisper_models = ["base", "small", "medium", "base.en"]


source_languages = {
"Arabic": "ar",
"Asturian ":"st",
"Belarusian":"be",
"Bulgarian":"bg",
"Czech":"cs",
"Danish":"da",
"German":"de",
"Greeek":"el",
"English":"en",
"Estonian":"et",
"Finnish":"fi",
"Swedish": "sv",
"Spanish":"es",
"Let the model analyze": "Let the model analyze"
}

DeepL_language_codes_for_translation = {
"Bulgarian": "BG",
"Czech": "CS",
"Danish": "DA",
"German": "DE",
"Greek": "EL",
"English": "EN",
"Spanish": "ES",
"Estonian": "ET",
"Finnish": "FI",
"French": "FR",
"Hungarian": "HU",
"Indonesian": "ID",
"Italian": "IT",
"Japanese": "JA",
"Lithuanian": "LT",
"Latvian": "LV",
"Dutch": "NL",
"Polish": "PL",
"Portuguese": "PT",
"Romanian": "RO",
"Russian": "RU",
"Slovak": "SK",
"Slovenian": "SL",
"Swedish": "SV",
"Turkish": "TR",
"Ukrainian": "UK",
"Chinese": "ZH"
}



transcribe_options = dict(beam_size=3, best_of=3, without_timestamps=False)


source_language_list = [key[0] for key in source_languages.items()]
translation_models_list = [key[0] for key in DeepL_language_codes_for_translation.items()]


device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("DEVICE IS: ")
print(device)
  
videos_out_path = Path("./videos_out")
videos_out_path.mkdir(parents=True, exist_ok=True)


def get_youtube(video_url):
    yt = YouTube(video_url)
    abs_video_path = yt.streams.filter(progressive=True, file_extension='mp4').order_by('resolution').desc().first().download()
    print("LADATATTU POLKUUN")
    print(abs_video_path)

    
    return abs_video_path

def speech_to_text(video_file_path, selected_source_lang, whisper_model):
    """
    # Youtube with translated subtitles using OpenAI Whisper and Opus-MT models.
    # Currently supports only English audio
    This space allows you to:
    1. Download youtube video with a given url
    2. Watch it in the first video component
    3. Run automatic speech recognition on the video using Whisper
    4. Translate the recognized transcriptions to Finnish, Swedish, Danish
    5. Burn the translations to the original video and watch the video in the 2nd video component
    
    Speech Recognition is based on OpenAI Whisper https://github.com/openai/whisper
    """
    
    if(video_file_path == None):
        raise ValueError("Error no video input")
    print(video_file_path)
    try:
        _,file_ending = os.path.splitext(f'{video_file_path}')
        print(f'file enging is {file_ending}')
        print("starting conversion to wav")
        os.system(f'ffmpeg -i "{video_file_path}" -ar 16000 -ac 1 -c:a pcm_s16le "{video_file_path.replace(file_ending, ".wav")}"')
        print("conversion to wav ready")



        print("starting whisper c++")
        srt_path = str(video_file_path.replace(file_ending, ".wav")) + ".srt"
        os.system(f'rm -f {srt_path}')
        if selected_source_lang == "Let the model analyze":
            os.system(f'./whisper.cpp/main "{video_file_path.replace(file_ending, ".wav")}" -t 4 -m ./whisper.cpp/models/ggml-{whisper_model}.bin -osrt')
        else:
            os.system(f'./whisper.cpp/main "{video_file_path.replace(file_ending, ".wav")}" -t 4 -l {source_languages.get(selected_source_lang)} -m ./whisper.cpp/models/ggml-{whisper_model}.bin -osrt')
        print("starting whisper done with whisper")
    except Exception as e:
        raise RuntimeError("Error converting video to audio")

    try:
        

        df = pd.DataFrame(columns = ['start','end','text'])
        srt_path = str(video_file_path.replace(file_ending, ".wav")) + ".srt"
        subs = pysrt.open(srt_path)


        objects = []
        for sub in subs:
            
            
            start_hours = str(str(sub.start.hours) + "00")[0:2] if len(str(sub.start.hours)) == 2 else str("0" + str(sub.start.hours) + "00")[0:2]
            end_hours = str(str(sub.end.hours) + "00")[0:2] if len(str(sub.end.hours)) == 2 else str("0" + str(sub.end.hours) + "00")[0:2]
            
            start_minutes = str(str(sub.start.minutes) + "00")[0:2] if len(str(sub.start.minutes)) == 2 else str("0" + str(sub.start.minutes) + "00")[0:2]
            end_minutes = str(str(sub.end.minutes) + "00")[0:2] if len(str(sub.end.minutes)) == 2 else str("0" + str(sub.end.minutes) + "00")[0:2]
            
            start_seconds = str(str(sub.start.seconds) + "00")[0:2] if len(str(sub.start.seconds)) == 2 else str("0" + str(sub.start.seconds) + "00")[0:2]
            end_seconds = str(str(sub.end.seconds) + "00")[0:2] if len(str(sub.end.seconds)) == 2 else str("0" + str(sub.end.seconds) + "00")[0:2]
            
            start_millis = str(str(sub.start.milliseconds) + "000")[0:3]
            end_millis = str(str(sub.end.milliseconds) + "000")[0:3]
            objects.append([sub.text, f'{start_hours}:{start_minutes}:{start_seconds}.{start_millis}', f'{end_hours}:{end_minutes}:{end_seconds}.{end_millis}'])

        for object in objects:
            srt_to_df = {
            'start': [object[1]],
            'end': [object[2]], 
            'text': [object[0]] 
            }
    
            df = pd.concat([df, pd.DataFrame(srt_to_df)])
        
                    
        return df
    
    except Exception as e:
        raise RuntimeError("Error Running inference with local model", e)


def translate_transcriptions(df, selected_translation_lang_2):
    if selected_translation_lang_2 is None:
            selected_translation_lang_2 = 'English'
    df.reset_index(inplace=True)
    
    print("start_translation")
    translations = []
    
    

    text_combined = ""
    for i, sentence in enumerate(df['text']):
        if i == 0:
            text_combined = sentence
        else:
            text_combined = text_combined + '\n' + sentence

    data = {'text': text_combined,
    'tag_spitting': 'xml',
    'target_lang': DeepL_language_codes.get(selected_source_lang_2)
           }
    response = requests.post('https://api-free.deepl.com/v2/translate', headers=headers, data=data)

    # Print the response from the server
    translated_sentences = json.loads(response.text)
    translated_sentences = translated_sentences['translations'][0]['text'].split('\n')
    df['translation'] = translated_sentences

    
    print("translations done")

    return (df)


def create_srt_and_burn(df, video_in):
    
    print("Starting creation of video wit srt")
    print("video in path is:")
    print(video_in)
    
       
    with open('testi.srt','w', encoding="utf-8") as file:
        for i in range(len(df)):
            file.write(str(i+1))
            file.write('\n')
            start = df.iloc[i]['start']
           
    

            file.write(f"{start}")
            
            stop = df.iloc[i]['end']
            
            
            file.write(' --> ')
            file.write(f"{stop}")
            file.write('\n')
            file.writelines(df.iloc[i]['translation'])
            if int(i) != len(df)-1:
                file.write('\n\n')
        
    print("SRT DONE")        
    try:
        file1 = open('./testi.srt', 'r', encoding="utf-8")
        Lines = file1.readlines()
        
        count = 0
        # Strips the newline character
        for line in Lines:
            count += 1
            print("{}".format(line))
                
        print(type(video_in))
        print(video_in)
        
        video_out = video_in.replace('.mp4', '_out.mp4')
        print("video_out_path")
        print(video_out)
        command = 'ffmpeg -i "{}" -y -vf subtitles=./testi.srt "{}"'.format(video_in, video_out)
        print(command)
        os.system(command)
        return video_out
    except Exception as e:
        print(e)
        return video_out


# ---- Gradio Layout -----
video_in = gr.Video(label="Video file", mirror_webcam=False)
youtube_url_in = gr.Textbox(label="Youtube url", lines=1, interactive=True)
video_out = gr.Video(label="Video Out", mirror_webcam=False)



df_init = pd.DataFrame(columns=['start','end','text'])
df_init_2 = pd.DataFrame(columns=['start','end','text','translation'])

selected_source_lang = gr.Dropdown(choices=source_language_list, type="value", value="Let the model analyze", label="Spoken language in video", interactive=True)
selected_translation_lang_2 = gr.Dropdown(choices=translation_models_list, type="value", value="English", label="In which language you want the transcriptions?", interactive=True)
selected_whisper_model = gr.Dropdown(choices=whisper_models, type="value", value="base", label="Selected Whisper model", interactive=True)

transcription_df = gr.DataFrame(value=df_init,label="Transcription dataframe", row_count=(0, "dynamic"), max_rows = 10, wrap=True, overflow_row_behaviour='paginate')
transcription_and_translation_df = gr.DataFrame(value=df_init_2,label="Transcription and translation dataframe", max_rows = 10, wrap=True, overflow_row_behaviour='paginate')


demo = gr.Blocks(css='''
#cut_btn, #reset_btn { align-self:stretch; }
#\\31 3 { max-width: 540px; }
.output-markdown {max-width: 65ch !important;}
''')
demo.encrypt = False
with demo:
    transcription_var = gr.Variable()
    
    with gr.Row():
        with gr.Column():
            gr.Markdown('''
            ### This space allows you to: 
            ##### 1. Download youtube video with a given URL
            ##### 2. Watch it in the first video component
            ##### 3. Run automatic speech recognition on the video using Whisper (Please remember to select translation language)
            ##### 4. Translate the recognized transcriptions to Finnish, Swedish, Danish
            ##### 5. Burn the translations to the original video and watch the video in the 2nd video component
            ''')
            
        with gr.Column():
            gr.Markdown('''
            ### 1. Insert Youtube URL below (Some examples below which I suggest to use for first tests)
            ##### 1. https://www.youtube.com/watch?v=nlMuHtV82q8&ab_channel=NothingforSale24
            ##### 2. https://www.youtube.com/watch?v=JzPfMbG1vrE&ab_channel=ExplainerVideosByLauren
            ##### 3. https://www.youtube.com/watch?v=S68vvV0kod8&ab_channel=Pearl-CohnTelevision
            ''')
            
    with gr.Row():
        with gr.Column():
            youtube_url_in.render()
            download_youtube_btn = gr.Button("Step 1. Download Youtube video")
            download_youtube_btn.click(get_youtube, [youtube_url_in], [
                video_in])
            print(video_in)
            

    with gr.Row():
        with gr.Column():
            video_in.render()
            with gr.Column():
                gr.Markdown('''
                ##### Here you can start the transcription and translation process.
                ##### Be aware that processing will last for a while (35 second video took around 20 seconds in my testing and might fail for longer videos)
                ''')
            selected_source_lang.render()
            selected_whisper_model.render()
            transcribe_btn = gr.Button("Step 2. Transcribe audio")
            transcribe_btn.click(speech_to_text, [video_in, selected_source_lang, selected_whisper_model], transcription_df)

            
    with gr.Row():
        gr.Markdown('''
        ##### Here you will get transcription  output
        ##### ''')

    with gr.Row():
        with gr.Column():
            transcription_df.render()
            
    with gr.Row():
        with gr.Column():
            gr.Markdown('''
            ##### Here you will get translated transcriptions. 
            ##### Please remember to select Spoken Language and wanted translation language
            ##### ''')
            selected_translation_lang_2.render()
            translate_transcriptions_button = gr.Button("Step 3. Translate transcription")
            translate_transcriptions_button.click(translate_transcriptions, [transcription_df, selected_translation_lang_2], transcription_and_translation_df)
            transcription_and_translation_df.render()
            
    with gr.Row():
        with gr.Column():
            gr.Markdown('''
            ##### Now press the Step 4. Button to create output video with translated transcriptions
            ##### ''')
            translate_and_make_srt_btn = gr.Button("Step 4. Create and burn srt to video")
            print(video_in)
            translate_and_make_srt_btn.click(create_srt_and_burn, [transcription_and_translation_df,video_in], [
                video_out])
            video_out.render()

                
demo.launch()