# Imports
import gradio as gr
import whisper
from pytube import YouTube
from transformers import pipeline, T5Tokenizer, T5ForConditionalGeneration, AutoTokenizer, AutoModelForSeq2SeqLM
from wordcloud import WordCloud
import re
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
self.summarizer = pipeline("summarization", model="facebook/bart-large-cnn")
# 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")
self.tokenizer = AutoTokenizer.from_pretrained("csebuetnlp/mT5_multilingual_XLSum")
self.model = AutoModelForSeq2SeqLM.from_pretrained("csebuetnlp/mT5_multilingual_XLSum")
def __call__(self, link, lang, size):
"""
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.
"""
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
# Transcribe the audio extracted from pytube
results = self.loaded_model.transcribe(path, language=lang)
# Perform summarization on the transcription
transcription_summary = self.summarizer(
results["text"], max_length=150, min_length=30, do_sample=False
)
#### Prueba
WHITESPACE_HANDLER = lambda k: re.sub('\s+', ' ', re.sub('\n+', ' ', k.strip()))
input_ids_sum = self.tokenizer(
[WHITESPACE_HANDLER(results["text"])],
return_tensors="pt",
padding="max_length",
truncation=True,
max_length=512
)["input_ids"]
output_ids_sum = self.model.generate(
input_ids=input_ids_sum,
max_length=130,
no_repeat_ngram_size=2,
num_beams=4
)[0]
summary = self.tokenizer.decode(
output_ids_sum,
skip_special_tokens=True,
clean_up_tokenization_spaces=False
)
#### Fin prueba
# 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()]
# Sentiment label
label = self.classifier(summary)[0]["label"]
# Generate WordCloud object
wordcloud = WordCloud().generate(results["text"])
# WordCloud image to display
wordcloud_image = wordcloud.to_image()
if lang == "english":
return (
results["text"],
transcription_summary[0]["summary_text"],
keywords,
label,
wordcloud_image,
)
else:
return (
results["text"],
summary,
keywords,
label,
wordcloud_image,
)
# return (
# results["text"],
# summary,
# keywords,
# label,
# wordcloud_image,
# )
def populate_metadata(self, link):
"""
Access to the YouTube video title and thumbnail image to further display it
params:
- link: a YouTube URL.
"""
self.yt = YouTube(link)
return self.yt.thumbnail_url, self.yt.title
def from_audio_input(self, lang, size, audio_file):
"""
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.
"""
if lang == "none":
lang = None
if size != self.current_size:
self.loaded_model = whisper.load_model(size)
self.current_size = size
results = self.loaded_model.transcribe(audio_file, language=lang)
# Perform summarization on the transcription
transcription_summary = self.summarizer(
results["text"], max_length=150, min_length=30, do_sample=False
)
#### Prueba
WHITESPACE_HANDLER = lambda k: re.sub('\s+', ' ', re.sub('\n+', ' ', k.strip()))
input_ids_sum = self.tokenizer(
[WHITESPACE_HANDLER(results["text"])],
return_tensors="pt",
padding="max_length",
truncation=True,
max_length=512
)["input_ids"]
output_ids_sum = self.model.generate(
input_ids=input_ids_sum,
max_length=130,
no_repeat_ngram_size=2,
num_beams=4
)[0]
summary = self.tokenizer.decode(
output_ids_sum,
skip_special_tokens=True,
clean_up_tokenization_spaces=False
)
#### Fin prueba
# 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()]
# Sentiment label
label = self.classifier(summary)[0]["label"]
# WordCloud object
wordcloud = WordCloud().generate(
results["text"]
)
wordcloud_image = wordcloud.to_image()
if lang == "english":
return (
results["text"],
transcription_summary[0]["summary_text"],
keywords,
label,
wordcloud_image,
)
else:
return (
results["text"],
summary,
keywords,
label,
wordcloud_image,
)
# return (
# results["text"],
# summary,
# keywords,
# label,
# wordcloud_image,
# )
gio = GradioInference()
title = "YouTube Insights"
description = "Your AI-powered video analytics tool"
block = gr.Blocks()
with block as demo:
gr.HTML(
"""
YouTube Insights 📹
Your AI-powered video analytics tool
"""
)
with gr.Group():
with gr.Tab("From YouTube"):
with gr.Box():
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"
)
link = gr.Textbox(
label="YouTube Link", placeholder="Enter YouTube link..."
)
title = gr.Label(label="Video Title")
with gr.Row().style(equal_height=True):
img = gr.Image(label="Thumbnail")
text = gr.Textbox(
label="Transcription",
placeholder="Transcription Output...",
lines=10,
).style(show_copy_button=True, container=True)
with gr.Row().style(equal_height=True):
summary = gr.Textbox(
label="Summary", placeholder="Summary Output...", lines=5
).style(show_copy_button=True, container=True)
keywords = gr.Textbox(
label="Keywords", placeholder="Keywords Output...", lines=5
).style(show_copy_button=True, container=True)
label = gr.Label(label="Sentiment Analysis")
wordcloud_image = gr.Image()
with gr.Row().style(equal_height=True):
clear = gr.ClearButton(
[link, title, img, text, summary, keywords, label, wordcloud_image], scale=1
)
btn = gr.Button("Get video insights", variant="primary", 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"):
with gr.Box():
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")
with gr.Row().style(equal_height=True):
text = gr.Textbox(
label="Transcription",
placeholder="Transcription Output...",
lines=10,
).style(show_copy_button=True, container=False)
with gr.Row().style(equal_height=True):
summary = gr.Textbox(
label="Summary", placeholder="Summary Output", lines=5
)
keywords = gr.Textbox(
label="Keywords", placeholder="Keywords Output", lines=5
)
label = gr.Label(label="Sentiment Analysis")
wordcloud_image = gr.Image()
with gr.Row().style(equal_height=True):
clear = gr.ClearButton([audio_file,text, summary, keywords, label, wordcloud_image], scale=1)
btn = gr.Button(
"Get video insights", variant="primary", scale=1
)
btn.click(
gio.from_audio_input,
inputs=[lang, size, audio_file],
outputs=[text, summary, keywords, label, wordcloud_image],
)
with block:
gr.Markdown("### Video Examples")
gr.Examples(["https://www.youtube.com/shorts/xDNzz8yAH7I","https://www.youtube.com/shorts/VrSJhTGMM90"], inputs=link)
gr.Markdown("### About the app:")
with gr.Accordion("What is YouTube Insights?", open=False):
gr.Markdown(
"YouTube Insights is a tool developed for academic purposes that allows you to analyze YouTube videos or audio files. It provides features like transcription, summarization, keyword extraction, sentiment analysis, and word cloud generation for multimedia content."
)
with gr.Accordion("How does YouTube Insights work?", open=False):
gr.Markdown(
"YouTube 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(
"YouTube 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."
)
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 YouTube Insights?" ,open=False):
gr.Markdown(
"YouTube 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)."
)
gr.HTML(
"""
2023 Master in Big Data & Data Science - Universidad Complutense de Madrid