import os
import requests
import json
import base64
os.system('git clone https://github.com/ggerganov/whisper.cpp.git')
os.system('git -C ./whisper.cpp fetch origin c10db6ea2883a4f77440fa8caeb296a0e351a58c')
os.system('git -C ./whisper.cpp reset --hard c10db6ea2883a4f77440fa8caeb296a0e351a58c')
os.system('make -C ./whisper.cpp')
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 large')
os.system('bash ./whisper.cpp/models/download-ggml-model.sh base.en')
import gradio as gr
from pathlib import Path
import pysrt
import pandas as pd
import re
import time
from pytube import YouTube
headers = {'Authorization': os.environ['DeepL_API_KEY']}
deepL_character_usage = 'UNKOWN'
import torch
whisper_models = ["base", "small", "medium", "large", "base.en"]
custom_models = ["belarus-small"]
combined_models = []
combined_models.extend(whisper_models)
combined_models.extend(custom_models)
#usage = requests.get('https://api.deepl.com/v2/usage', headers=headers)
#usage = json.loads(usage.text)
#deepL_character_usage = str(usage['character_count'])
#print("deepL_character_usage")
LANGUAGES = {
"en": "English",
"zh": "Chinese",
"de": "German",
"es": "Spanish",
"ru": "Russian",
"ko": "Korean",
"fr": "French",
"ja": "Japanese",
"pt": "Portuguese",
"tr": "Turkish",
"pl": "Polish",
"ca": "Catalan",
"nl": "Dutch",
"ar": "Arabic",
"sv": "Swedish",
"it": "Italian",
"id": "Indonesian",
"hi": "Hindi",
"fi": "Finnish",
"vi": "Vietnamese",
"he": "Hebrew",
"uk": "Ukrainian",
"el": "Greek",
"ms": "Malay",
"cs": "Czech",
"ro": "Romanian",
"da": "Danish",
"hu": "Hungarian",
"ta": "Tamil",
"no": "Norwegian",
"th": "Thai",
"ur": "Urdu",
"hr": "Croatian",
"bg": "Bulgarian",
"lt": "Lithuanian",
"la": "Latin",
"mi": "Maori",
"ml": "Malayalam",
"cy": "Welsh",
"sk": "Slovak",
"te": "Telugu",
"fa": "Persian",
"lv": "Latvian",
"bn": "Bengali",
"sr": "Serbian",
"az": "Azerbaijani",
"sl": "Slovenian",
"kn": "Kannada",
"et": "Estonian",
"mk": "Macedonian",
"br": "Breton",
"eu": "Basque",
"is": "Icelandic",
"hy": "Armenian",
"ne": "Nepali",
"mn": "Mongolian",
"bs": "Bosnian",
"kk": "Kazakh",
"sq": "Albanian",
"sw": "Swahili",
"gl": "Galician",
"mr": "Marathi",
"pa": "Punjabi",
"si": "Sinhala",
"km": "Khmer",
"sn": "Shona",
"yo": "Yoruba",
"so": "Somali",
"af": "Afrikaans",
"oc": "Occitan",
"ka": "Georgian",
"be": "Belarusian",
"tg": "Tajik",
"sd": "Sindhi",
"gu": "Gujarati",
"am": "Amharic",
"yi": "Yiddish",
"lo": "Lao",
"uz": "Uzbek",
"fo": "Faroese",
"ht": "Haitian creole",
"ps": "Pashto",
"tk": "Turkmen",
"nn": "Nynorsk",
"mt": "Maltese",
"sa": "Sanskrit",
"lb": "Luxembourgish",
"my": "Myanmar",
"bo": "Tibetan",
"tl": "Tagalog",
"mg": "Malagasy",
"as": "Assamese",
"tt": "Tatar",
"haw": "Hawaiian",
"ln": "Lingala",
"ha": "Hausa",
"ba": "Bashkir",
"jw": "Javanese",
"su": "Sundanese",
}
# language code lookup by name, with a few language aliases
source_languages = {
**{language: code for code, language in LANGUAGES.items()},
"Burmese": "my",
"Valencian": "ca",
"Flemish": "nl",
"Haitian": "ht",
"Letzeburgesch": "lb",
"Pushto": "ps",
"Panjabi": "pa",
"Moldavian": "ro",
"Moldovan": "ro",
"Sinhalese": "si",
"Castilian": "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):
"""
# PLEASE NOTE CURRENTLY TRANSLATIONS DO NOT WORK BECAUSE OF ISSUES WITH DEEPL
# 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 fast Whisper models
4. Translate the recognized transcriptions to 26 languages supported by deepL (If free API usage for the month is not yet fully consumed)
5. Download generated subtitles in .vtt and .srt formats
6. Watch the the original video with generated subtitles
Speech Recognition is based on models from OpenAI Whisper https://github.com/openai/whisper
This space is using c++ implementation by https://github.com/ggerganov/whisper.cpp
"""
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")
except Exception as e:
raise RuntimeError("Error Running inference with local model", e)
try:
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 -l "auto" -m ./whisper.cpp/models/ggml-{whisper_model}.bin -osrt')
else:
if whisper_model in custom_models:
os.system(f'./whisper.cpp/main "{video_file_path.replace(file_ending, ".wav")}" -t 4 -l {source_languages.get(selected_source_lang)} -m ./converted_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 running Whisper cpp model")
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)])
except Exception as e:
print("Error creating srt df")
# try:
# #usage = requests.get('https://api-free.deepl.com/v2/usage', headers=headers)
# #usage = json.loads(usage.text)
# #char_count = str(usage['character_count'])
# #print('Usage is at: ' + str(usage['character_count']) + ' characters')
# #if usage['character_count'] >= 490000:
# # print("USAGE CLOSE TO LIMIT")
# except Exception as e:
# print('Error with DeepL API requesting usage count')
return df
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_for_translation.get(selected_translation_lang_2)
}
# try:
# #usage = requests.get('https://api-free.deepl.com/v2/usage', headers=headers)
# #usage = json.loads(usage.text)
# #deepL_character_usage = str(usage['character_count'])
# #try:
# # print('Usage is at: ' + deepL_character_usage + 'characters')
# #except Exception as e:
# # print(e)
# #
# #if int(deepL_character_usage) <= 490000:
# # print("STILL CHARACTERS LEFT")
# # 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
# else:
# df['translation'] = df['text']
# except Exception as e:
# print("EXCEPTION WITH DEEPL API")
# print(e)
df['translation'] = df['text']
print("translations done")
print("Starting SRT-file creation")
print(df.head())
df.reset_index(inplace=True)
with open('subtitles.vtt','w', encoding="utf-8") as file:
print("Starting WEBVTT-file creation")
for i in range(len(df)):
if i == 0:
file.write('WEBVTT')
file.write('\n')
else:
file.write(str(i+1))
file.write('\n')
start = df.iloc[i]['start']
file.write(f"{start.strip()}")
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("WEBVTT DONE")
with open('subtitles.srt','w', encoding="utf-8") as file:
print("Starting SRT-file creation")
for i in range(len(df)):
file.write(str(i+1))
file.write('\n')
start = df.iloc[i]['start']
file.write(f"{start.strip()}")
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")
subtitle_files = ['subtitles.vtt','subtitles.srt']
return df, subtitle_files
# def burn_srt_to_video(srt_file, video_in):
# print("Starting creation of video wit srt")
# try:
# video_out = video_in.replace('.mp4', '_out.mp4')
# print(os.system('ls -lrth'))
# print(video_in)
# print(video_out)
# command = 'ffmpeg -i "{}" -y -vf subtitles=./subtitles.srt "{}"'.format(video_in, video_out)
# os.system(command)
# return video_out
# except Exception as e:
# print(e)
# return video_out
def create_video_player(subtitle_files, video_in):
with open(video_in, "rb") as file:
video_base64 = base64.b64encode(file.read())
with open('./subtitles.vtt', "rb") as file:
subtitle_base64 = base64.b64encode(file.read())
video_player = f'''
'''
#video_player = gr.HTML(video_player)
return video_player
# ---- 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', '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,label="Transcription and translation dataframe", max_rows = 10, wrap=True, overflow_row_behaviour='paginate')
subtitle_files = gr.File(
label="Download srt-file",
file_count="multiple",
type="file",
interactive=False,
)
video_player = gr.HTML('
video will be played here after you press the button at step 4')
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 fast Whisper models
4. Translate the recognized transcriptions to 26 languages supported by deepL
5. Download generated subtitles in .vtt and .srt formats
6. Watch the the original video with generated subtitles
''')
with gr.Column():
gr.Markdown('''
### 1. Copy any non-private Youtube video URL to box below or click one of the examples.
(But please **consider using short videos** so others won't get queued)
Then press button "1. Download Youtube video"-button:
''')
examples = gr.Examples(examples=
[ "https://www.youtube.com/watch?v=fLeJJPxua3E&ab_channel=Motiversity",
"https://www.youtube.com/watch?v=JzPfMbG1vrE&ab_channel=ExplainerVideosByLauren",
"https://www.youtube.com/watch?v=S68vvV0kod8&ab_channel=Pearl-CohnTelevision"],
label="Examples", inputs=[youtube_url_in])
# Inspiration from https://huggingface.co/spaces/vumichien/whisper-speaker-diarization
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 some time. With base model it is around 3x speed
**Please select source language** for better transcriptions. Using 'Let the model analyze' makes mistakes sometimes and may lead to bad transcriptions
''')
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('''
### PLEASE READ BELOW
### DEEPL TRANSLATIONS DO NOT CURRENTLY WORK
### Because of big demand for this demo all credits for translation might be gone already for the month. In this case we return the original transcript from this component :(
### I might make some adjustments in the future for the translation to use some other model if all the API credits have been used but this is the situation for now
### Translation credits will reset every 5th of month.
Here you will can translate transcriptions to 26 languages.
If spoken language is not in the list, translation might not work. In this case original transcriptions are used.
''')
gr.Markdown(f'''
DEEPL DOES NOT WORK CURRENTLY, Sorry for inconvenience
DeepL API character usage:
{deepL_character_usage if deepL_character_usage is not None else ''}/500 000 characters
If usage is over 490 000 characters original transcriptions will be used for subtitles. This value might not properly update so if you get transcriptions in original language that might be the reason.
API usage resets on 5th of every month.
''')
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, subtitle_files])
transcription_and_translation_df.render()
with gr.Row():
with gr.Column():
gr.Markdown('''##### From here you can download subtitles in .srt or .vtt format''')
subtitle_files.render()
with gr.Row():
with gr.Column():
gr.Markdown('''
##### Now press the Step 4. Button to create output video with translated transcriptions
##### ''')
create_video_button = gr.Button("Step 4. Create and add subtitles to video")
print(video_in)
create_video_button.click(create_video_player, [subtitle_files,video_in], [
video_player])
video_player.render()
demo.launch()