import whisper
from pytubefix import YouTube
from pytubefix.cli import on_progress
import requests
import time
import streamlit as st
from streamlit_lottie import st_lottie
import numpy as np
import os
from typing import Iterator
from io import StringIO
from utils import write_vtt, write_srt
import ffmpeg
from languages import LANGUAGES
import torch
from zipfile import ZipFile
from io import BytesIO
import base64
import pathlib
import re
st.set_page_config(page_title="Auto Subtitled Video Generator", page_icon=":movie_camera:", layout="wide")
torch.cuda.is_available()
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
# Model options: tiny, base, small, medium, large
loaded_model = whisper.load_model("small", device=DEVICE)
current_size = "None"
# Define a function that we can use to load lottie files from a link.
def load_lottieurl(url: str):
r = requests.get(url)
if r.status_code != 200:
return None
return r.json()
APP_DIR = pathlib.Path(__file__).parent.absolute()
LOCAL_DIR = APP_DIR / "local_youtube"
LOCAL_DIR.mkdir(exist_ok=True)
save_dir = LOCAL_DIR / "output"
save_dir.mkdir(exist_ok=True)
col1, col2 = st.columns([1, 3])
with col1:
lottie = load_lottieurl("https://assets8.lottiefiles.com/packages/lf20_jh9gfdye.json")
st_lottie(lottie)
with col2:
st.write("""
## Auto Subtitled Video Generator
##### Input a YouTube video link and get a video with subtitles.
###### ➠ If you want to transcribe the video in its original language, select the task as "Transcribe"
###### ➠ If you want to translate the subtitles to English, select the task as "Translate"
###### I recommend starting with the base model and then experimenting with the larger models, the small and medium models often work well. """)
def download_video(link):
yt = YouTube(link, use_oauth=True, on_progress_callback=on_progress)
ys = yt.streams.get_highest_resolution()
video = ys.download(filename=f"{save_dir}/youtube_video.mp4")
return video
def convert(seconds):
return time.strftime("%H:%M:%S", time.gmtime(seconds))
def change_model(current_size, size):
if current_size != size:
loaded_model = whisper.load_model(size)
return loaded_model
else:
raise Exception("Model size is the same as the current size.")
def inference(link, loaded_model, task):
yt = YouTube(link, use_oauth=True, on_progress_callback=on_progress)
ys = yt.streams.get_audio_only()
path = ys.download(filename=f"{save_dir}/audio.mp3", mp3=True)
if task == "Transcribe":
options = dict(task="transcribe", best_of=5)
results = loaded_model.transcribe(path, **options)
vtt = getSubs(results["segments"], "vtt", 80)
srt = getSubs(results["segments"], "srt", 80)
lang = results["language"]
return results["text"], vtt, srt, lang
elif task == "Translate":
options = dict(task="translate", best_of=5)
results = loaded_model.transcribe(path, **options)
vtt = getSubs(results["segments"], "vtt", 80)
srt = getSubs(results["segments"], "srt", 80)
lang = results["language"]
return results["text"], vtt, srt, lang
else:
raise ValueError("Task not supported")
def getSubs(segments: Iterator[dict], format: str, maxLineWidth: int) -> str:
segmentStream = StringIO()
if format == 'vtt':
write_vtt(segments, file=segmentStream, maxLineWidth=maxLineWidth)
elif format == 'srt':
write_srt(segments, file=segmentStream, maxLineWidth=maxLineWidth)
else:
raise Exception("Unknown format " + format)
segmentStream.seek(0)
return segmentStream.read()
def get_language_code(language):
if language in LANGUAGES.keys():
detected_language = LANGUAGES[language]
return detected_language
else:
raise ValueError("Language not supported")
def generate_subtitled_video(video, audio, transcript):
video_file = ffmpeg.input(video)
audio_file = ffmpeg.input(audio)
ffmpeg.concat(video_file.filter("subtitles", transcript), audio_file, v=1, a=1).output("youtube_sub.mp4").run(quiet=True, overwrite_output=True)
video_with_subs = open("youtube_sub.mp4", "rb")
return video_with_subs
def main():
size = st.selectbox("Select Model Size (The larger the model, the more accurate the transcription will be, but it will take longer)", ["tiny", "base", "small", "medium", "large-v3"], index=1)
loaded_model = change_model(current_size, size)
st.write(f"Model is {'multilingual' if loaded_model.is_multilingual else 'English-only'} "
f"and has {sum(np.prod(p.shape) for p in loaded_model.parameters()):,} parameters.")
link = st.text_input("YouTube Link (The longer the video, the longer the processing time)", placeholder="Input YouTube link and press enter")
task = st.selectbox("Select Task", ["Transcribe", "Translate"], index=0)
if task == "Transcribe":
if st.button("Transcribe"):
with st.spinner("Transcribing the video..."):
results = inference(link, loaded_model, task)
video = download_video(link)
lang = results[3]
detected_language = get_language_code(lang)
col3, col4 = st.columns(2)
with col3:
st.video(video)
# Split result["text"] on !,? and . , but save the punctuation
sentences = re.split("([!?.])", results[0])
# Join the punctuation back to the sentences
sentences = ["".join(i) for i in zip(sentences[0::2], sentences[1::2])]
text = "\n\n".join(sentences)
with open("transcript.txt", "w+", encoding='utf8') as f:
f.writelines(text)
f.close()
with open(os.path.join(os.getcwd(), "transcript.txt"), "rb") as f:
datatxt = f.read()
with open("transcript.vtt", "w+",encoding='utf8') as f:
f.writelines(results[1])
f.close()
with open(os.path.join(os.getcwd(), "transcript.vtt"), "rb") as f:
datavtt = f.read()
with open("transcript.srt", "w+",encoding='utf8') as f:
f.writelines(results[2])
f.close()
with open(os.path.join(os.getcwd(), "transcript.srt"), "rb") as f:
datasrt = f.read()
with col4:
with st.spinner("Generating Subtitled Video"):
video_with_subs = generate_subtitled_video(video, f"{save_dir}/audio.mp3", "transcript.srt")
st.video(video_with_subs)
st.balloons()
zipObj = ZipFile("YouTube_transcripts_and_video.zip", "w")
zipObj.write("transcript.txt")
zipObj.write("transcript.vtt")
zipObj.write("transcript.srt")
zipObj.write("youtube_sub.mp4")
zipObj.close()
ZipfileDotZip = "YouTube_transcripts_and_video.zip"
with open(ZipfileDotZip, "rb") as f:
datazip = f.read()
b64 = base64.b64encode(datazip).decode()
href = f"\
Download Transcripts and Video\
"
st.markdown(href, unsafe_allow_html=True)
elif task == "Translate":
if st.button("Translate to English"):
with st.spinner("Translating to English..."):
results = inference(link, loaded_model, task)
video = download_video(link)
lang = results[3]
detected_language = get_language_code(lang)
col3, col4 = st.columns(2)
with col3:
st.video(video)
# Split result["text"] on !,? and . , but save the punctuation
sentences = re.split("([!?.])", results[0])
# Join the punctuation back to the sentences
sentences = ["".join(i) for i in zip(sentences[0::2], sentences[1::2])]
text = "\n\n".join(sentences)
with open("transcript.txt", "w+", encoding='utf8') as f:
f.writelines(text)
f.close()
with open(os.path.join(os.getcwd(), "transcript.txt"), "rb") as f:
datatxt = f.read()
with open("transcript.vtt", "w+",encoding='utf8') as f:
f.writelines(results[1])
f.close()
with open(os.path.join(os.getcwd(), "transcript.vtt"), "rb") as f:
datavtt = f.read()
with open("transcript.srt", "w+",encoding='utf8') as f:
f.writelines(results[2])
f.close()
with open(os.path.join(os.getcwd(), "transcript.srt"), "rb") as f:
datasrt = f.read()
with col4:
with st.spinner("Generating Subtitled Video"):
video_with_subs = generate_subtitled_video(video, f"{save_dir}/audio.mp3", "transcript.srt")
st.video(video_with_subs)
st.balloons()
zipObj = ZipFile("YouTube_transcripts_and_video.zip", "w")
zipObj.write("transcript.txt")
zipObj.write("transcript.vtt")
zipObj.write("transcript.srt")
zipObj.write("youtube_sub.mp4")
zipObj.close()
ZipfileDotZip = "YouTube_transcripts_and_video.zip"
with open(ZipfileDotZip, "rb") as f:
datazip = f.read()
b64 = base64.b64encode(datazip).decode()
href = f"\
Download Transcripts and Video\
"
st.markdown(href, unsafe_allow_html=True)
else:
st.info("Please select a task.")
if __name__ == "__main__":
main()