import streamlit as st
import requests
import base64
import os
from moviepy.editor import VideoFileClip
from pytube import YouTube
from youtube_transcript_api import YouTubeTranscriptApi
from youtube_transcript_api._errors import CouldNotRetrieveTranscript
import whisper
import ffmpeg
import re
import tempfile
from huggingface_hub import InferenceClient
st.set_page_config(layout="wide", initial_sidebar_state="collapsed")
PROMPT = """Act as the author and provide a comprehensive detailed article in the same language as the transcript
in markdown format that has a H1 main title(example "# <this is a title> ") and broken down into H2 subtitles (example "## <this is a title> ") for the following transcript
You must follow the rules:
- Write the article in markdown format
- Create a main title for the article as markdown H1 and break the article into subtitles where each subtitle is markdown H2
- Article must be in the same language as the transcript
- summary should be informative and act as a replacement for the original transcript to the point that the user doesn't have to go back to read the transcript
- Summary should not mention the author or speaker at all should act as your independent writing without referencing the original transcript or speaker.
- You can use bullet points within the article
Transcript:
{} \n\n Article:"""
LLM = {
"llama3-8b": {'prompt': f"""<|begin_of_text|><|start_header_id|>user<|end_header_id|>
{PROMPT}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
""", 'endpoint': "meta-llama/Meta-Llama-3-8B-Instruct"}
}
@st.cache_resource()
def load_whisper(model):
return whisper.load_model(model)
@st.cache_data
def download_video(url):
if "youtube" in url or "youtu.be" in url:
yt = YouTube(url)
video = yt.streams.get_highest_resolution()
filename = video.download()
else:
response = requests.get(url, stream=True)
filename = url.split("/")[-1]
with open(filename, "wb") as file:
for chunk in response.iter_content(chunk_size=1024):
if chunk:
file.write(chunk)
return filename
@st.cache_data
def convert_to_audio(video_filename):
video = VideoFileClip(video_filename)
audio_filename = video_filename.replace(".mp4", ".mp3")
audio = video.audio
audio.write_audiofile(audio_filename, codec="mp3")
return audio_filename
@st.cache_data
def summarise(prompt, llm):
model = InferenceClient(LLM[llm]["endpoint"])
user_message = LLM[llm]["prompt"].format(prompt)
return model.text_generation(user_message, max_new_tokens=1024)
def delete_files(video_filename, audio_filename):
delete_file(video_filename)
delete_file(audio_filename)
def delete_file(filename):
if os.path.exists(filename):
os.remove(filename)
st.info(f"File '{os.path.basename(filename)}' deleted from the server.")
@st.cache_data
def transcribe_whisper(_model, audio_filepath):
return _model.transcribe(audio_filepath)["text"]
def get_media_download_link(media_type, file_path):
with open(file_path, "rb") as file:
contents = file.read()
encoded = base64.b64encode(contents).decode("utf-8")
media_href = f"data:file/{media_type};base64,{encoded}"
st.markdown(
f'<a href="{media_href}" download="{os.path.basename(file_path)}">Download {os.path.basename(file_path)}</a>',
unsafe_allow_html=True,
)
@st.cache_data
def generate_summaries(_summarizer, text, min_length=50, max_length=500):
paragraphs = text.split("\n\n")
summaries = []
for paragraph in paragraphs:
summary = _summarizer(
paragraph, max_length=max_length, min_length=min_length, do_sample=False
)
summaries.append(summary[0]["summary_text"].strip())
return "\n\n".join(summaries)
def main():
st.title("VidScripter")
st.write("#### A One Stop Solution to Video Transcription")
c1, c2 = st.columns(2)
c1.write(
"""
- Enter the video URL in the text input box.
- Click the **Fetch** button to fetch the video.
- Once the video is fetched, you can perform the following actions:
- Fetch transcript from YouTube API (if available) by clicking the **Fetch Transcript** button.
- Transcribe the video using the Whisper model by clicking the **Transcribe (Whisper)** button.
- The transcript will be displayed in a text area below.
- A summary of the transcript will also be generated by the selected LLM.
- The summary will be displayed in a text area below.
- You can download the video, audio, transcript or summary by clicking the respective download buttons.
"""
)
whisper_model = load_whisper("base")
url = c2.text_input("Enter the video URL")
llm = c2.selectbox("Select LLM", list(LLM.keys()), index=0)
fetch_button = c2.button("Fetch")
st.session_state.setdefault("load_state", False)
if fetch_button or st.session_state.load_state:
st.session_state.load_state = True
if url:
process_video(url, whisper_model, llm)
def process_video(url, whisper_model, llm):
yt = YouTube(url)
video_id = yt.video_id
try:
video_filename = download_video(url)
st.success("Video fetched successfully")
except Exception:
video_filename = None
st.warning("Video could not be fetched")
try:
audio_filename = (
convert_to_audio(video_filename) if video_filename is not None else None
)
if video_filename is not None:
st.success("Audio converted successfully")
else:
st.info("No Video to convert into Audio")
except Exception:
audio_filename = None
st.warning("Audio coud not be converted")
text_filename = (
os.path.basename(video_filename).replace(".mp4", ".txt")
if video_filename is not None
else "transcript.txt"
)
emp = st.empty()
col1, col2, col3, col4 = st.columns(4)
if "youtube" in url or "youtu.be" in url:
process_youtube_video(video_id, col3, emp, text_filename, llm)
process_whisper_transcript(whisper_model, audio_filename, col4, text_filename)
with col1:
if video_filename is not None and st.button("Download Video"):
with st.spinner("Encoding Video"):
get_media_download_link("video", video_filename)
with col2:
if audio_filename is not None and st.button("Download Audio"):
with st.spinner("Encoding Audio"):
get_media_download_link("audio", audio_filename)
def process_youtube_video(video_id, col, emp, text_filename, llm):
try:
transcript_list = YouTubeTranscriptApi.list_transcripts(video_id)
transcripts = [transcript for transcript in transcript_list]
if transcripts:
transcript_options = {
f"{transcript.language} ({transcript.language_code})": transcript
for transcript in transcripts
}
transcript_option = emp.selectbox(
"Select a transcript", list(transcript_options.keys())
)
selected_transcript = transcript_options[transcript_option]
st.session_state.setdefault("api_transcript", False)
if col.button("Fetch Transcript") or st.session_state.api_transcript:
st.session_state.api_transcript = True
transcript_text = selected_transcript.fetch()
transcript_text = "\n".join(
[re.sub("\s+", " ", chunk["text"]) for chunk in transcript_text]
)
c1, c2 = st.columns(2)
with c1:
modified_text = st.text_area(
"Transcript", transcript_text, height=500
)
st.download_button("Download Transcript", modified_text, text_filename)
with c2:
openai_summarization = summarise(
modified_text, llm
)
summarized_text = st.text_area(
"Summarized Transcript", openai_summarization, height=500
)
st.download_button("Download Summary", summarized_text, text_filename)
except CouldNotRetrieveTranscript:
emp.warning("Could Not Retrieve API Transcripts for this video.")
except Exception as e:
emp.warning(f"Error Fetching API Transcripts for this video. {e}")
def process_whisper_transcript(whisper_model, audio_filename, col, text_filename):
if audio_filename is not None:
st.session_state.setdefault("whisper_transcript", False)
if col.button("Transcribe (Whisper)") or st.session_state.whisper_transcript:
st.session_state.whisper_transcript = True
whisper_text = transcribe_whisper(whisper_model, audio_filename)
modified_text = st.text_area("Transcript", whisper_text, height=500)
st.download_button("Download", modified_text, text_filename)
if __name__ == "__main__":
main()