E3S Web of Conferences 399, 04015 (2023) https://doi.org/10.

1051/e3sconf/202339904015
ICONNECT-2023

VIDEO TRANSCRIPT SUMMARIZER

Ilampiray P1, Naveen Raju D1, Thilagavathy A1, Mohamed Tharik M1, Madhan Kishore
S1, Nithin A.S1, Infant Raj I2
1Department of Computer Science and Engineering, R.M.K. Engineering College Thiruvallur, Tamil
Nadu, India.
2Department of Computer Science and Engineering, K.Ramakrishnan College of Engineering, Trichy

- 621112, Tamil Nadu, India.

ABSTRACT. In today’s world, a large number of videos are

uploaded in everyday, which contains information about something.
The major challenge is to find the right video and understand the
correct content, because there are lot of videos available some videos
will contain useless content and even though the perfect content
available that content should be required to us. If we not found right
one it wastes your full effort and full time to extract the correct
usefull information. We propose an innovation idea which uses NLP
processing for text extraction and BERT Summarization for Text
Summarization. This provides a video main content in text
description and abstractive summary, enabling users to discriminate
between relevant and irrelevant information according to their needs.
Furthermore, our experiments show that the joint model can attain
good results with informative, concise, and readable multi-line video
description and summary in a human evaluation.

1 Introduction
NLP, a subset of Artificial Intelligence, is a field that concentrates on the interaction
between machines and human languages. Its primary objective is to enable machines to
understand, interpret, and generate natural language text or speech.. Video summarization,
which involves generating concise and accurate summaries of longer videos, is an important
application of NLP. The goal is to produce short and coherent summaries that capture the key
information of the original video. This technology can be valuable in situations where time
is limited, or when there is a need for a quick overview of the video content. Video
summarization typically involves a combination of techniques such as text extraction, audio
analysis, and image processing, among others. Our main idea is to able to find the short
summary of YouTube video and present it in a textual format. As short summarization of text
methods will very helpful for users because we can read the important content in short period
of time. In the field of NLP we generate summaries of transcripts and produce human-
readable outputs. Nowadays from child to older people are easy to YouTube video for many

 Corresponding author: ilampiray.vp@gmail.com

© The Authors, published by EDP Sciences. This is an open access article distributed under the terms of the Creative Commons
Attribution License 4.0 (https://creativecommons.org/licenses/by/4.0/).
E3S Web of Conferences 399, 04015 (2023) https://doi.org/10.1051/e3sconf/202339904015
ICONNECT-2023

purposes like educational, entertainment and many other kinds of genre, it is very necessary
to find out exact content we required. In the Internet we can see that so many are videos are
long but not even contain good useful information which wastes our time. We can directly
move to the main to content of YouTube video by removing the useless information of the
videos. The main goal of our this paper is to increase work efficiency and save the time of
a user. There are many users just seeing the thumbnail of the Youtube video and catchy title
turns eager into the video, wastes the time by watching useless information. Students often
search for YouTube videos before exams, but due to time constraints, they may watch them
at double speed, which can lead to confusion about the subject matter. Having access to
recorded sessions and transcripts of meetings can be helpful in obtaining a summary of the
video content, saving time and effort. The main focus of our paper is to extract the most
important information from the transcript and present it in a concise paragraph. Our objective
is to save users' time by providing them with relevant and useful information on their desired
topic. Automatic text summarization techniques have been developed in recent years to
facilitate this process. Text summarization involves transforming text into a condensed
version that conveys the main message to users. Although text summarization is a challenging
task due to the limitations of machines in understanding human language and knowledge, it
offers various benefits such as content organization, summarization, data retrieval, and
question answering. While earlier studies primarily focused on simplifying and summarizing
single document texts, advancements in technology have paved the way for more efficient
and rapid text summarization methods.

2 Literature survey
The research paper introduced Latent Dirichlet Allocation (LDA) as an effective technique
for document summarization. Their proposed LDA summarizing model consists of three
stages [1]. In the initial phase, the subtitle file undergoes pre-processing, including the
removal of stop words and other tasks. The LDA model is then trained using the subtitles in
the second step to generate a list of keywords, which will be utilized to extract relevant
sentences. Finally, in the third phase, the summary is generated based on the extracted
keywords. Comparatively, the quality of the summaries produced by the LDA-based approach
surpasses that of TF-IDF and LSA summaries. The paper also presented Stream Hover, a
platform designed for explaining and summarizing transcripts of live streamed videos [2].
They explored a neural extractive summarization model that learns vector representations of
audio files and extracts significant observations from subtitles. These observations are utilized
to construct summaries using a vector quantized autoencoder. Additionally, the paper
proposed a system that can generate subtitles for movies in English, Hindi, or Malayalam,
depending on user preference. The system comprises three components: audio extraction,
voice recognition, and subtitle generation. For audio extraction, the FFMPEG platform is
employed to convert audio files of any format into .wav (Waveform Audio) format [3]. The
.wav file obtained from the audio extraction process is utilized to generate subtitles in the
form of a .srt file. The content of the audio, obtained through the Google Translate API, is
processed for speech attention. In the Subtitle creation module, the synchronized lyrics from
the .srt file are combined with the video using the Moviepy video enhancement library in
Python [4]. The research paper introduces a system that generates abstractive summaries for
videos covering various topics such as cooking, cuisine, software configuration, and sports.
To expand the vocabulary, the model is pre-trained on large English datasets using transfer
learning. Transcript pre-processing is also performed to improve sentence structure and
punctuation in ASR system results [5]. The evaluation of the results on the How2 and
WikiHow datasets involves the use of ROUGE and Content-F1 scoring metrics. The paper
categorizes different Text Summarization Methods, with a focus on giving more importance

2
E3S Web of Conferences 399, 04015 (2023) https://doi.org/10.1051/e3sconf/202339904015
ICONNECT-2023

to abstractive text summarization [6]. The authors express their belief that abstractive
summarization, despite being more challenging and computationally intensive than extractive
summarization, holds greater potential for generating more natural and human-like
summaries. This suggests that there may be further advancements in this field, offering new
perspectives from computational, cognitive, and linguistic standpoints. The study
recommends the ASoVS model, a hybrid end-to-end approach, for generating video
descriptions and text summaries in an abstractive manner [7]. The model incorporates a deep
neural network and captures various aspects of the video, such as people's traits (gender, age,
emotion), scenes, objects, and behaviors, to provide a multi-line description. The utilization
of OCR technology enables the conversion of images into text. The paper outlines the
strategies employed for subtitle generation, which involves three modules: Audio Extraction,
responsible for converting MPEG-compliant input files to .wav format; Speech Recognition,
which utilizes Hidden Markov Models (HMMs) to recognize extracted speech using language
and acoustic models; and Subtitle Generation, which produces synchronized .txt/.srt files.
This approach is particularly beneficial for individuals who are deaf, have reading difficulties,
or are learning to read [8]. The study also presents a multilingual speech-to-text conversion
method that involves feeding human voice utterances into a Speech-To-Text (STT) system,
utilizing Mel-Frequency Cepstral Coefficient feature extraction, Minimum Distance
Classifier, and Support Vector Machine techniques for voice classification [9]. The evaluation
of text summaries in the paper is done using ROUGE Metrics (Recall-Oriented Understudy
for Gisting Evaluation), which compare the computer-generated summary with human-
written ideal summaries by measuring the overlapping units such as n-grams, word pairs, and
word sequences [10]. The systems discussed in the papers [11] are not applicable to videos
that lack readily available subtitles. Additionally, these systems are limited to processing
English videos only. Furthermore, the absence of a media player in the proposed system [11]
requires the entire video to be uploaded for subtitle generation. In contrast, the system
described in the work [12] generates video descriptions solely based on the visual content,
without considering the audio. This system is particularly suitable for generating descriptions
for CCTV footages. However, it should be noted that the system in [12] is designed for audio
files and is not compatible with video files. The system presented in the article [3] exclusively
works with text input, lacking the capability to extract subtitles from videos or generate
subtitles for videos. In the work [13], the focus is solely on translation rather than text
summarization. The generated output in this system is the translated version of the text
obtained through speech recognition. Moreover, the paper [14] only addresses the extraction
of subtitles from videos, without providing a summarized version of the text. The vast amount
of information available on the internet can be overwhelming and emotionally draining for
individuals. To address this issue, summaries are employed to condense texts into a more
concise form while retaining the essential information. The main goal of a summary is to
effectively convey the key information in a concise manner. However, generating a useful
summary can be challenging, particularly when the original document contains repetitive
sentences. In such cases, reducing the document size would result in a loss of content [15].
This challenge is known as automatic text summarization, which aims to create a compact
and effective summary that captures the crucial information and overall meaning of the
original document. Automated text summarization is a complex task as computers lack the
linguistic understanding and knowledge possessed by humans. Consequently, automated text
summarization requires advanced technologies and techniques and is a time-consuming
process. Nonetheless, despite the challenges involved, text summarization is gaining
importance in fields such as journalism, research, and content creation, where the ability to
extract key information quickly is vital.

3
E3S Web of Conferences 399, 04015 (2023) https://doi.org/10.1051/e3sconf/202339904015
ICONNECT-2023

3 Methodology
This paper majorly focus on providing clean, clear and correct summary of the YouTube
videos that users don’t need to waste their time at. This paper use more popular python
libraries each for each purpose. They are YouTube Transcipt api for transcript extraction,
BERT is for summarization library it is combination of GPT and BART, google translate api
for translation and Flask framework is for Backend Connection in Fig.1.

Fig. 1. System Architecture

3.1 Getting URL

A Uniform Resource Locator (URL) is a reference to an internet resource that provides the
location of that resource on a computer network. It serves as a means to retrieve information
from the internet. When a user enters a YouTube URL in the search box, the system processes
the URL in the background to obtain the necessary information. Then check whether it is
valid URL or NOT, If it is valid URL then shortened the link process will happened in the
backend. Then required YouTube video will get from the link. It will process for the Video
To Text Extraction. In this user will enter the URL, we will check whether the URL is valid
or not. If the URL is invalid it will return the error message. If the URL is valid, then it will
pass correct URL in next step that it will pass the URL to NLP model.

3.2 Video to text extraction

After the required YouTube video is attained, then we will pass the URL from NLP model
file to utubeextract.py file. Then by using the youtube_transcript_api we will get the
transcript for required YouTube video. There are three different ways to transcript one is

4
E3S Web of Conferences 399, 04015 (2023) https://doi.org/10.1051/e3sconf/202339904015
ICONNECT-2023

automatically transcript generated, second-one is manually transcript generated and third-

one is videos that doesn’t contain transcript. Videos without transcripts are being excluded
from our selection. Once we obtain the transcript of a video using an API, we proceed to
process the text for subsequent procedures. This involves removing commas, punctuation
marks, and full stops, as they are essential in determining sentence boundaries. This process
can done by using the python library “punctuator. Then our next step is to apply the text
preprocessing method for extracted transcript. This above method is like removing
punctuations, stopword, exclamatory expression, stemming and for scanning purposes. The
purpose of above task is to shortening a wide length of paragraph into a short summary. That
summary will contain important information in the video. In NLP text summarization we tend
to use the Extractive Summarization. This Extractive summarization delivers the summary
that contains only the important phrases and sentences. The site will continuously monitor the
data in the user given web page and if there is any change in the amount specified, an
immediate alert is passed to the user or customer. Once the user provides the required
information, our website will scrap the required data from the product‘s web page. For
scrapping the data our website uses beautiful soup module. It is a Python module, which is
basically used to scrap the data from website in Fig.2.

Fig. 2. Text Extraction

3.3 NLP processing

This Process is for Text Summarization. NLP contains many method for text summarization.
We will talk about the text summarization is process of shortening the big paragraph into
short summary. If the paragraph contains lots of lines, we need more time to cover the
content, we have time scarcity so we want only a main report of that text. We can able to
convert the large text into to small text by removing unimportant information. The process
of breaking down lengthy text into digestible paragraphs or sentences is known as NLP Text
Summarization. This above process will retrieve the important information in Fig.3.

5
E3S Web of Conferences 399, 04015 (2023) https://doi.org/10.1051/e3sconf/202339904015
ICONNECT-2023

Fig. 3. NLP Processing

3.4 BERT summarizer

BERT is Bidirectional Encoder Representations from Transformers. This BERT introduces

new advanced approach to deliver NLP tasks. BERT is one of the important algorithm of
Natural Language Processing Models. The important information can be retained and
extracting the relevant information by Extractive Text Summarization. This Extractive
Summarization is more challenging. In this progress we using superior embeddings that
provided by encoder models like BERT. By taking two supervised approaches and use the
BERT sentence embeddings to build an extractive summarizer. The first considers only
embeddings and their derivatives. Based on the internal structure of the article we can able to
good summarizer can parse meaning and select correct meaningful sentences. The
Unsupervised Text-Rank model is the baseline approach. Other than approaches are
incorporates the sequential information and advantage of well known particular to new
corpuses. In fact, many publishers have been deployed this strategy. Baseline for the second
approach is Lead. Rouge-1 and Rouge-L F1 metric are important things in supervised models
outperform in Fig.4.

Fig. 4. BERT Summarization Diagram

3.5 Language translation

A language translator is a very handy application that helps us to communicate in different

languages by translating our language to the desired language. In earlier times, when there
were no Language Translation Applications, it was very difficult for people to communicate
with people coming from different parts of the world. In this we can create our own language
translation project using python. Our objective is to create a Language Translator which
would help us translate a word, sentence or even a paragraph to another language. We will
try to incorporate as many languages as possible. We will be using T-kinter Module to build
our GUI for the project and google-trans library to present us with a number of languages

6
E3S Web of Conferences 399, 04015 (2023) https://doi.org/10.1051/e3sconf/202339904015
ICONNECT-2023

that are a part of it.

4. Results and discussions

Initially the verifying the valid URL or not, if valid we will check whether the transcript is
available or not. Then we will go for fetching the required YouTube video and using
YouTube transcript api we will fetch the transcript of YouTube video. Videos lacking
transcripts are being filtered out. Once we acquire the video transcript through an API, we
proceed to process the text for subsequent steps. This involves eliminating commas,
punctuation marks, and full stops, which are crucial for identifying sentence boundaries. This
process can done by using the python library “punctuator. Then our next step is to apply the
text preprocessing method for extracted transcript in Fig. 5.

Fig. 5. Block Diagram of Proposed System

This paper is used for following functions
Crash-Course: Students can able to find out a particular topic of the subject in the you-tube
video and precisely able to get the quick read of the youtube video.
Education in the online era has paved the way for generating summaries and allowing
students to create their own notes based on video class transcripts. Similarly, in video and
audio conferences, transcripts can be used to outline discussions held during team
conferences. In the context of patent research, the extraction of crucial claims from patents
and research papers can be facilitated using figures such as Fig. 6, Fig. 7, and Fig. 8.

7
E3S Web of Conferences 399, 04015 (2023) https://doi.org/10.1051/e3sconf/202339904015
ICONNECT-2023

Fig. 6. Extractive Summarizer

Fig. 7. Metric Table of Our Results

Fig. 8. Metric Table of Processing time and memory Usage

8
E3S Web of Conferences 399, 04015 (2023) https://doi.org/10.1051/e3sconf/202339904015
ICONNECT-2023

4.1 Screenshots

Fig. 9. Home Page (In this Home page, the user paste URL in video URL box, The URL which is
copied from the YouTube video)

Fig. 10. Home page (after paste the link, in backend required video will process and it takes
some time)

Fig. 11. Output Page (here the summarized text in English, It also display the comparison
between word count of before and after summarization)

9
E3S Web of Conferences 399, 04015 (2023) https://doi.org/10.1051/e3sconf/202339904015
ICONNECT-2023

Fig. 12. Here we can able to download summarized text and save it as text file.

5 Future works
In the future, we intend to extend our work in the based of extensions. This extension need
to be available in all the browser, social media, and all other Video website like(YouTube,
Share-Chat etc). User only purpose is need to add the extension and need to select the required
video It will automatically fetch the link and display the required shortened summary.
browser and machine learning and artificial intelligence technology.

6 Conclusions
In conclusion, our website can save time for the user .Instaed of Seeing the whole buffer
waste content of the video we will prefix see the what is main content of YouTube Video
and Know which video will perfectly for us. It save the time and effort of the user. By using
our website their burden will be reduced for search the right YouTube Video. Our website
will also provide Multi-Language Summarization and made the availability to Text-to-
Speech Process also. We are confident that our paper will effectively address the needs of
users by saving their time and efforts. Our approach aims to provide users with only the
relevant and useful information on the topics that interest them, eliminating the need to watch
lengthy videos. This time saved can be utilized for further knowledge acquisition and
exploration.

REFERENCES
1 Alrumiah, S. S., Al-Shargabi, A. A. Educational Videos Subtitles’ Summarization Using
Latent Dirichlet Allocation and Length Enhancement. CMC-Computers, Materials &
Continua, 70, 3 (2022).
2 Sangwoo Cho, Franck Dernoncourt , Tim Ganter, Trung Bui, Nedim Lipka, Walter
Chang, Hailin Jin, Jonathan Brandt, Hassan Foroosh, Fei Liu, “StreamHover: Livestream
Transcript Summarization and Annotation”, (2022).

10
E3S Web of Conferences 399, 04015 (2023) https://doi.org/10.1051/e3sconf/202339904015
ICONNECT-2023

3 S. Chopra, M. Auli, and A. M. Rush, ‘‘Abstractive sentence summarization with attentive

recurrent neural networks,’’ in Proc. Conf. North Amer. Chapter Assoc. Comput.
Linguistics Hum. Lang. Technol., (2016).
4 Ghadage, Yogita H. and Sushama Shelke. “Speech to text conversion for multilingual
languages.” 2016 International Conference on Communication and Signal Processing
(ICCSP) (2016).
5 Pravin Khandare, Sanket Gaikwad, Aditya Kukade, Rohit Panicker, Swaraj Thamke,
“Audio Data Summarization system using Natural Language Processing, ” International
Research Journal of Engineering and Technology (IRJET), 6, 9 (2019).
6 S.M.Mahedy Hasan, Md. Fazle Rabbi, Arifa Islam Champa, Md. Asif Zaman, “An
Effective Diabetes Prediction System Using Machine Learning Techniques”, 2nd
International Conference on Advanced Information and Communication Technology
(ICAICT), (2020).
7 Prof. S. A. Aher , Hajari Ashwini M , Hase Megha S, Jadhav Snehal B, Pawar Snehal S,
“Generating Subtitles Automatically For Sound in Videos, ” International Journal of
Modern Trends in Engineering and Research (IJMTER), 3, 3 (2016).
8 Aiswarya K R, “ Automatic Multiple Language Subtitle Generation for Videos, ”
International Research Journal of Engineering and Technology (IRJET), 7, 5 (2020).
9 Savelieva, Alexandra & Au-Yeung, Bryan & Ramani, Vasanth. Abstractive
Summarization of Spoken and Written Instructions with BERT (2020).
10 Patil, S. et al. “Multilingual Speech and Text Recognition and Translation using Image.”
International journal of engineering research and technology 5 (2016).
11 S. Sah, S. Kulhare, A. Gray, S. Venugopalan, E. Prud'Hom meaux and R. Ptucha,
"Semantic Text Summarization of Long Videos," IEEE Winter Conference on
Applications of Computer Vision (WACV), (2017).
12 A. Dilawari and M. U. G. Khan, "ASoVS: Abstractive Summarization of Video
Sequences," in IEEE Access, 7 (2019).
13 Lin, Chin-Yew, “ROUGE: A Package for Automatic Evaluation of Summaries,” In
Proceedings of 2004, Association for Computational Linguistics, Barcelona, Spain.
14 A. E. B. Ajmal and R. P. Haroon, “Maximal marginal relevance based malayalam text
summarization with successive thresholds,” International Journal on Cybernetics and
Informatics, 5, 2 (2016).
15 M. Allahyari et al., “Text summarization techniques: A brief survey,” International
Journal of Advanced Computer Science and Applications, 8, 10 (2017).

11