# Libraries
import gradio as gr
import whisper
from pytube import YouTube
from transformers import pipeline, T5Tokenizer, T5ForConditionalGeneration, AutoTokenizer, AutoModelForSeq2SeqLM
import torch
from wordcloud import WordCloud
import re
import os
class GradioInference:
def __init__(self):
# OpenAI's Whisper model sizes
self.sizes = list(whisper._MODELS.keys())
# Whisper's available languages for ASR
self.langs = ["none"] + sorted(list(whisper.tokenizer.LANGUAGES.values()))
# Default size
self.current_size = "base"
# Default model size
self.loaded_model = whisper.load_model(self.current_size)
# Initialize Pytube Object
self.yt = None
# Initialize summary model for English
self.bart_summarizer = pipeline("summarization", model="facebook/bart-large-cnn", truncation=True)
# Initialize Multilingual summary model
self.mt5_tokenizer = AutoTokenizer.from_pretrained("csebuetnlp/mT5_multilingual_XLSum", truncation=True)
self.mt5_model = AutoModelForSeq2SeqLM.from_pretrained("csebuetnlp/mT5_multilingual_XLSum")
# Initialize VoiceLabT5 model and tokenizer
self.keyword_model = T5ForConditionalGeneration.from_pretrained(
"Voicelab/vlt5-base-keywords"
)
self.keyword_tokenizer = T5Tokenizer.from_pretrained(
"Voicelab/vlt5-base-keywords"
)
# Sentiment Classifier
self.classifier = pipeline("text-classification", model="lxyuan/distilbert-base-multilingual-cased-sentiments-student", return_all_scores=False)
def __call__(self, link, lang, size, progress=gr.Progress()):
"""
Call the Gradio Inference python class.
This class gets access to a YouTube video using python's library Pytube and downloads its audio.
Then it uses the Whisper model to perform Automatic Speech Recognition (i.e Speech-to-Text).
Once the function has the transcription of the video it proccess it to obtain:
- Summary: using Facebook's BART transformer.
- KeyWords: using VoiceLabT5 keyword extractor.
- Sentiment Analysis: using Hugging Face's default sentiment classifier
- WordCloud: using the wordcloud python library.
"""
try:
progress(0, desc="Starting analysis")
if self.yt is None:
self.yt = YouTube(link)
# Pytube library to access to YouTube audio stream
path = self.yt.streams.filter(only_audio=True)[0].download(filename="tmp.mp4")
if lang == "none":
lang = None
if size != self.current_size:
self.loaded_model = whisper.load_model(size)
self.current_size = size
progress(0.20, desc="Transcribing")
# Transcribe the audio extracted from pytube
results = self.loaded_model.transcribe(path, language=lang)
progress(0.40, desc="Summarizing")
# Perform summarization on the transcription
transcription_summary = self.bart_summarizer(
results["text"],
max_length=150,
min_length=30,
do_sample=False,
truncation=True
)
# Multilingual summary with mt5
WHITESPACE_HANDLER = lambda k: re.sub('\s+', ' ', re.sub('\n+', ' ', k.strip()))
input_ids_sum = self.mt5_tokenizer(
[WHITESPACE_HANDLER(results["text"])],
return_tensors="pt",
padding="max_length",
truncation=True,
max_length=512
)["input_ids"]
output_ids_sum = self.mt5_model.generate(
input_ids=input_ids_sum,
max_length=256,
no_repeat_ngram_size=2,
num_beams=4
)[0]
summary = self.mt5_tokenizer.decode(
output_ids_sum,
skip_special_tokens=True,
clean_up_tokenization_spaces=False
)
# End multilingual summary
progress(0.60, desc="Extracting Keywords")
# Extract keywords using VoiceLabT5
task_prefix = "Keywords: "
input_sequence = task_prefix + results["text"]
input_ids = self.keyword_tokenizer(
input_sequence,
return_tensors="pt",
truncation=False
).input_ids
output = self.keyword_model.generate(
input_ids,
no_repeat_ngram_size=3,
num_beams=4
)
predicted = self.keyword_tokenizer.decode(output[0], skip_special_tokens=True)
keywords = [x.strip() for x in predicted.split(",") if x.strip()]
formatted_keywords = "\n".join([f"β’ {keyword}" for keyword in keywords])
progress(0.80, desc="Extracting Sentiment")
# Define a dictionary to map labels to emojis
sentiment_emojis = {
"positive": "Positive ππΌ",
"negative": "Negative ππΌ",
"neutral": "Neutral πΆ",
}
# Sentiment label
label = self.classifier(summary)[0]["label"]
# Format the label with emojis
formatted_sentiment = sentiment_emojis.get(label, label)
progress(0.90, desc="Generating Wordcloud")
# Generate WordCloud object
wordcloud = WordCloud(colormap = "Oranges").generate(results["text"])
# WordCloud image to display
wordcloud_image = wordcloud.to_image()
if lang == "english" or lang == "none":
return (
results["text"],
transcription_summary[0]["summary_text"],
formatted_keywords,
formatted_sentiment,
wordcloud_image,
)
else:
return (
results["text"],
summary,
formatted_keywords,
formatted_sentiment,
wordcloud_image,
)
except:
gr.Error(message="Restricted Content. Choose a different video")
finally:
gr.Info("Success!")
def populate_metadata(self, link):
"""
Access to the YouTube video title and thumbnail image to further display it
params:
- link: a YouTube URL.
"""
if not link:
return None, None
self.yt = YouTube(link)
return self.yt.thumbnail_url, self.yt.title
def from_audio_input(self, lang, size, audio_file, progress=gr.Progress()):
"""
Call the Gradio Inference python class.
Uses it directly the Whisper model to perform Automatic Speech Recognition (i.e Speech-to-Text).
Once the function has the transcription of the video it proccess it to obtain:
- Summary: using Facebook's BART transformer.
- KeyWords: using VoiceLabT5 keyword extractor.
- Sentiment Analysis: using Hugging Face's default sentiment classifier
- WordCloud: using the wordcloud python library.
"""
try:
progress(0, desc="Starting analysis")
if lang == "none":
lang = None
if size != self.current_size:
self.loaded_model = whisper.load_model(size)
self.current_size = size
progress(0.20, desc="Transcribing")
results = self.loaded_model.transcribe(audio_file, language=lang)
progress(0.40, desc="Summarizing")
# Perform summarization on the transcription
transcription_summary = self.bart_summarizer(
results["text"], max_length=150, min_length=30, do_sample=False, truncation=True
)
# Multilingual summary with mt5
WHITESPACE_HANDLER = lambda k: re.sub('\s+', ' ', re.sub('\n+', ' ', k.strip()))
input_ids_sum = self.mt5_tokenizer(
[WHITESPACE_HANDLER(results["text"])],
return_tensors="pt",
padding="max_length",
truncation=True,
max_length=512
)["input_ids"]
output_ids_sum = self.mt5_model.generate(
input_ids=input_ids_sum,
max_length=130,
no_repeat_ngram_size=2,
num_beams=4
)[0]
summary = self.mt5_tokenizer.decode(
output_ids_sum,
skip_special_tokens=True,
clean_up_tokenization_spaces=False
)
# End multilingual summary
progress(0.60, desc="Extracting Keywords")
# Extract keywords using VoiceLabT5
task_prefix = "Keywords: "
input_sequence = task_prefix + results["text"]
input_ids = self.keyword_tokenizer(
input_sequence,
return_tensors="pt",
truncation=False
).input_ids
output = self.keyword_model.generate(
input_ids,
no_repeat_ngram_size=3,
num_beams=4
)
predicted = self.keyword_tokenizer.decode(output[0], skip_special_tokens=True)
keywords = [x.strip() for x in predicted.split(",") if x.strip()]
formatted_keywords = "\n".join([f"β’ {keyword}" for keyword in keywords])
progress(0.80, desc="Extracting Sentiment")
# Define a dictionary to map labels to emojis
sentiment_emojis = {
"positive": "Positive ππΌ",
"negative": "Negative ππΌ",
"neutral": "Neutral πΆ",
}
# Sentiment label
label = self.classifier(summary)[0]["label"]
# Format the label with emojis
formatted_sentiment = sentiment_emojis.get(label, label)
progress(0.90, desc="Generating Wordcloud")
# WordCloud object
wordcloud = WordCloud(colormap = "Oranges").generate(
results["text"]
)
wordcloud_image = wordcloud.to_image()
if lang == "english" or lang == "none":
return (
results["text"],
transcription_summary[0]["summary_text"],
formatted_keywords,
formatted_sentiment,
wordcloud_image,
)
else:
return (
results["text"],
summary,
formatted_keywords,
formatted_sentiment,
wordcloud_image,
)
except:
gr.Error(message="Exceeded audio size. Choose a different audio")
finally:
gr.Info("Success!")
def from_article(self, article, progress=gr.Progress()):
"""
Call the Gradio Inference python class.
Acepts the user's text imput, then it performs:
- Summary: using Facebook's BART transformer.
- KeyWords: using VoiceLabT5 keyword extractor.
- Sentiment Analysis: using Hugging Face's default sentiment classifier
- WordCloud: using the wordcloud python library.
"""
try:
progress(0, desc="Starting analysis")
progress(0.30, desc="Summarizing")
# Perform summarization on the transcription
transcription_summary = self.bart_summarizer(
article, max_length=150, min_length=30, do_sample=False, truncation=True
)
# Multilingual summary with mt5
WHITESPACE_HANDLER = lambda k: re.sub('\s+', ' ', re.sub('\n+', ' ', k.strip()))
input_ids_sum = self.mt5_tokenizer(
[WHITESPACE_HANDLER(article)],
return_tensors="pt",
padding="max_length",
truncation=True,
max_length=512
)["input_ids"]
output_ids_sum = self.mt5_model.generate(
input_ids=input_ids_sum,
max_length=130,
no_repeat_ngram_size=2,
num_beams=4
)[0]
summary = self.mt5_tokenizer.decode(
output_ids_sum,
skip_special_tokens=True,
clean_up_tokenization_spaces=False
)
# End multilingual summary
progress(0.60, desc="Extracting Keywords")
# Extract keywords using VoiceLabT5
task_prefix = "Keywords: "
input_sequence = task_prefix + article
input_ids = self.keyword_tokenizer(
input_sequence,
return_tensors="pt",
truncation=False
).input_ids
output = self.keyword_model.generate(
input_ids,
no_repeat_ngram_size=3,
num_beams=4
)
predicted = self.keyword_tokenizer.decode(output[0], skip_special_tokens=True)
keywords = [x.strip() for x in predicted.split(",") if x.strip()]
formatted_keywords = "\n".join([f"β’ {keyword}" for keyword in keywords])
progress(0.80, desc="Extracting Sentiment")
# Define a dictionary to map labels to emojis
sentiment_emojis = {
"positive": "Positive ππΌ",
"negative": "Negative ππΌ",
"neutral": "Neutral πΆ",
}
# Sentiment label
label = self.classifier(summary)[0]["label"]
# Format the label with emojis
formatted_sentiment = sentiment_emojis.get(label, label)
progress(0.90, desc="Generating Wordcloud")
# WordCloud object
wordcloud = WordCloud(colormap = "Oranges").generate(
article
)
wordcloud_image = wordcloud.to_image()
return (
transcription_summary[0]["summary_text"],
formatted_keywords,
formatted_sentiment,
wordcloud_image,
)
except:
gr.Error(message="Exceeded text size. Choose a different audio")
finally:
gr.Info("Success!")
gio = GradioInference()
title = "Media Insights"
description = "Your AI-powered video analytics tool"
theme = gr.themes.Soft(spacing_size="md", radius_size="md")
block = gr.Blocks(theme=theme)
# Gradio Interface
with block as demo:
# Title
gr.HTML(
"""
MEDIA INSIGHTS π‘
Your AI-powered media analytics tool β¨
"""
)
# Group of tabs
with gr.Group():
with gr.Tab("From YouTube πΉ"):
with gr.Box():
# Model Size and Language selections
with gr.Row().style(equal_height=True):
size = gr.Dropdown(
label="Speech-to-text Model Size", choices=gio.sizes, value="base"
)
lang = gr.Dropdown(
label="Language (Optional)", choices=gio.langs, value="none"
)
link = gr.Textbox(
label="YouTube Link", placeholder="Enter YouTube link..."
)
# Video Metadata
with gr.Row().style(equal_height=True):
with gr.Column(variant="panel", scale=1):
title = gr.Label(label="Video Title")
img = gr.Image(label="Thumbnail").style(height=350)
# Video Transcription
with gr.Column(variant="panel", scale=1):
text = gr.Textbox(
label="Transcription",
placeholder="Transcription Output...",
lines=18,
).style(show_copy_button=True)
# Video block of summary, keywords , sent. analysis and wordcloud
with gr.Row().style(equal_height=True):
summary = gr.Textbox(
label="Summary", placeholder="Summary Output...", lines=5
).style(show_copy_button=True)
keywords = gr.Textbox(
label="Keywords", placeholder="Keywords Output...", lines=5
).style(show_copy_button=True)
label = gr.Label(label="Sentiment Analysis")
wordcloud_image = gr.Image(label="WordCloud")
# Buttons
with gr.Row():
btn = gr.Button("Get Video Insights π", variant="primary", scale=1)
clear = gr.ClearButton(
[link, title, img, text, summary, keywords, label, wordcloud_image],
value="Clear ποΈ", scale=1
)
btn.click(
gio,
inputs=[link, lang, size],
outputs=[text, summary, keywords, label, wordcloud_image],
)
link.change(gio.populate_metadata, inputs=[link], outputs=[img, title])
with gr.Tab("From Audio file ποΈ"):
with gr.Box():
# Model selections
with gr.Row().style(equal_height=True):
size = gr.Dropdown(
label="Model Size", choices=gio.sizes, value="base"
)
lang = gr.Dropdown(
label="Language (Optional)", choices=gio.langs, value="none"
)
audio_file = gr.Audio(type="filepath")
# Audio transcription
with gr.Row().style(equal_height=True):
text = gr.Textbox(
label="Transcription",
placeholder="Transcription Output...",
lines=10,
).style(show_copy_button=True)
# Audio analysis
with gr.Row().style(equal_height=True):
summary = gr.Textbox(
label="Summary", placeholder="Summary Output...", lines=5
).style(show_copy_button=True)
keywords = gr.Textbox(
label="Keywords", placeholder="Keywords Output...", lines=5
).style(show_copy_button=True)
label = gr.Label(label="Sentiment Analysis")
wordcloud_image = gr.Image(label="WordCloud")
with gr.Row():
btn = gr.Button(
"Get Audio Insights π", variant="primary"
)
clear = gr.ClearButton([audio_file,text, summary, keywords, label, wordcloud_image], value="Clear ποΈ")
btn.click(
gio.from_audio_input,
inputs=[lang, size, audio_file],
outputs=[text, summary, keywords, label, wordcloud_image],
)
with gr.Tab("From Article π"):
with gr.Box():
# Text input from user
with gr.Row().style(equal_height=True):
article = gr.Textbox(
label="Text",
placeholder="Paste your text...",
lines=10,
).style(show_copy_button=True)
# Text analysis
with gr.Row().style(equal_height=True):
summary = gr.Textbox(
label="Summary", placeholder="Summary Output...", lines=5
).style(show_copy_button=True)
keywords = gr.Textbox(
label="Keywords", placeholder="Keywords Output...", lines=5
).style(show_copy_button=True)
label = gr.Label(label="Sentiment Analysis")
wordcloud_image = gr.Image(label="WordCloud")
with gr.Row():
btn = gr.Button(
"Get Text insights π", variant="primary")
clear = gr.ClearButton([article, summary, keywords, label, wordcloud_image], value="Clear ποΈ")
btn.click(
gio.from_article,
inputs=[article],
outputs=[summary, keywords, label, wordcloud_image],
)
# Open text example
with open(os.path.join(os.path.dirname(__file__), "texts/India_Canada.txt"), "r") as file:
text_example_content = file.read()
with block:
# Video Examples
gr.Markdown("### Video Examples")
gr.Examples(["https://www.youtube.com/shorts/xDNzz8yAH7I",
"https://youtu.be/MnrJzXM7a6o",
"https://youtu.be/FKjj1tNcbtM"],
inputs=link)
# Audio Examples
gr.Markdown("### Audio Examples")
gr.Examples([[os.path.join(os.path.dirname(__file__),"audios/EnglishLecture.mp4")]], inputs=audio_file)
# Text Examples
gr.Markdown("### Text Examples")
with gr.Accordion("News text example", open=False):
gr.Examples([[text_example_content]], inputs=article)
# FAQs section
gr.Markdown("### About the app:")
with gr.Accordion("What is Media Insights?", open=False):
gr.Markdown(
"Media Insights is a tool developed for academic purposes that allows you to analyze YouTube videos, audio files or some text. It provides features like transcription, summarization, keyword extraction, sentiment analysis, and word cloud generation for multimedia content."
)
with gr.Accordion("How does Media Insights work?", open=False):
gr.Markdown(
"Media Insights leverages several powerful AI models and libraries. It uses OpenAI's Whisper for Automatic Speech Recognition (ASR) to transcribe audio content. It summarizes the transcribed text using Facebook's BART model, extracts keywords with VoiceLabT5, performs sentiment analysis with DistilBERT, and generates word clouds."
)
with gr.Accordion("What languages are supported for the analysis?", open=False):
gr.Markdown(
"Media Insights supports multiple languages for transcription and analysis. You can select your preferred language from the available options when using the app."
)
with gr.Accordion("Can I analyze audio files instead of YouTube videos?", open=False):
gr.Markdown(
"Yes, you can analyze audio files directly. Simply upload your audio file to the app, and it will provide the same transcription, summarization, keyword extraction, sentiment analysis, and word cloud generation features. In addition, you can also paste your article or text of your preference, to get all the insights directly from it."
)
with gr.Accordion("What are the different model sizes available for transcription?", open=False):
gr.Markdown(
"The app uses a Speech-to-text model that has different training sizes, from tiny to large. Hence, the bigger the model the accurate the transcription."
)
with gr.Accordion("How long does it take to analyze a video or audio file?", open=False):
gr.Markdown(
"The time taken for analysis may vary based on the duration of the video or audio file and the selected model size. Shorter content will be processed more quickly."
)
with gr.Accordion("Who developed Media Insights?" ,open=False):
gr.Markdown(
"Media Insights was developed by students as part of the 2022/23 Master's in Big Data & Data Science program at Universidad Complutense de Madrid for academic purposes (Trabajo de Fin de Master)."
)
# Page footer
gr.HTML(
"""
Trabajo de Fin de MΓ‘ster - Grupo 3
2022/23 Master in Big Data & Data Science - Universidad Complutense de Madrid