Mynah: An Assistive Tool for the Speech Impaired

Abstract

[evdate 2021-05-15] Mynah: An assistive tool for the speech impaired

Details of Authors Author #1 Full Name BHATT, Umakant Datt Student ID

School SENG Department CPEG Major CPEG Year of study 4 Email

Mobile No.

Author #2 Full Name SHRESTHA, Aayush Student ID

School SENG Department CPEG Major CPEG Year of study 4 Email

Mobile No.

Author #3 Full Name SUNG, Ka Hong Student ID

School SENG Department CPEG Major CPEG Year of study 4 Email

Mobile No.

Advised by Dr. Desmond Yau-chat TSOI

1. Introduction 1.1 Abstract Communication plays a vital role in all aspects of our life including professional to personal life and everything within it. The quintessence of communication is the ability to establish commonness with the other party. This is important for conveying, imparting and exchanging thoughts, emotions, feelings and ideas. In particular, most of the time humans use visual and auditory channels to understand each. According to the World Health Organisation in 2015, the deaf and mute population had reached over 70 million [1]. Such individuals face difficulties in verbally expressing themselves, making their integration with the other members of the society difficult and ultimately hindering their development. With such a high proportion of the population facing difficulties in expressing themselves, we propose Mynah, an assistive tool for the speech impaired individuals.

Current methods for the speech impaired includes: assistance of a hand sign expert, and writing texts. Although hand signs and text are popularly used as a means to communicate among the speech impaired individuals, the former requires extensive special education to understand different signs and the latter requires effort and is inefficient while communicating back and forth. Moveover, hiring a hand sign expert is not scalable and is financially taxing. Furthermore, in situations such as discussions and presentations, communicating solely through text makes the task of conveying thoughts and ideas through emotion difficult.

Moreover

With the limitations posed by the current methods in mind, Mynah provides an effective but affordable solution by leveraging the current advances on deep generative modeling with minimal effort to produce realistic talking head model videos. Mynah takes short samples of speech (10 sentences), which could be theirs if less impaired or anyone close to them if severely impaired, to train a custom voice model and a picture whose face the user wants to model initially as a one-off input. After both the inputs are collected, the users will only be required to select the model and type a text script to generate talking head model video. As opposed to other alternatives which require external assistance or require individuals to carry an additional material to write on, Mynah operates as a mobile application which can be accessed anywhere and at any time conveniently. Moreover, Mynah generates both auditory and visual outputs through the means of talking head model video which provides a more effective approach for communication.

Recent advances in deep learning modeling have led to several solutions which generate similar head model videos using short samples of input. Although the end goal of producing head model videos remains the same, most of the work related to head model generation is through taking speech as a source of input, which is not suitable for our case. By combining the techniques used in several state-of-the-art works, we have developed Mynah, which takes in text as an input to

produce realistic head models with the goal of helping the speech impaired population. With the aim to help the disadvantaged speech impaired population integrate smoothly in the society by providing a means to easily communicate, we aim to design and develop Mynah by improving and optimizing various techniques and customizing them to meet these goals.

1.2 Objectives The primary objective of this application is to create a system that reduces the inconvenience of speech-impaired people so that they can have normal conversations, discussions and exchange with ordinary people in real life with this application, ultimately allowing them to integrate into the society. This system can be separated into the following four objectives: 1. Produce customized speech using a few samples of audio input of the user or any other speaker depending on the severity of the speech impairment. 2. Generate a realistic talking video for mute people to have smooth and effective communication with ordinary people. 3. Eliminate the need for external assistance with a convenient system that automatically produces realistic head models using a text input system. 4. Design and develop an easily operated mobile application interface for the speech impaired population to generate talking head model videos.

2. Methodology 2.1 Design 2.1.1 Application

Figure 1: Overall view of our application development.

To eliminate the need for external assistance with a convenient text input system for the speech impaired, we have decided to build a mobile application for users to access our system. Figure 1 outlines the basic flow of our application. In the mobile application, it will first perform data collection. Users can themselves or let others who can speak, depending on the severity, read the displayed 10 sentences on the screen, which will be requested for one time only and used for training a custom voice. An alternative for users is choosing a trained voice model among all available trained models. After the optional training is done, users can start to send the text that they want to say with a face image. After the mobile application receives a text from users, it will request the backend through the Fetch Application Programming Interface (API). The backend will receive the request through the webserver. The text with the face image provided by the client is used by the pre-trained model to generate a realistic head model video which is given as a response back to the client. The realistic head model video is displayed in the mobile application for mute people to have a visualized communication tool. It is also useful in creating presentations and discussion videos using the realistic talking head video for speech-impaired users to express themselves. 2.1.2 System Design

Figure 2: Proposed neural network architecture with three main components: text-to-speech, speech-to-landmarks, and landmarks-to-video Figure 2 shows an overview of the proposed neural network architecture. It accepts a still image, a short recording of the user’s voice, and the script in text format as input. First, the system will train the Real-time Voice Cloning model with the extracted user’s voice. After finishing the training with a few training steps, the text will be processed by the network and produce the user-specified speech. Inside the speech-to-landmarks section, the generated speech will be used for producing the

general face landmarks by the landmark generation module. Then, the realistic head model mapping will take the still image and the generated face landmarks as input and generate the talking face video. Finally, we obtain the realistic talking head video by concatenating the user-specified voice and the face movement video. In practice, once the system is finished training and ready to use, the user can send a request and the text will be processed by the system. The system will produce a realistic talking video in a few seconds and enables users to have fast communication with ordinary people without any communication medium problem. 2.1.2.1 Speech Module

Figure 3: Our proposed neural network architecture for speech module

Architecture in Figure 3 consists of three models, speaker encoder, synthesizer and the vocoder. The speaker encoder model is used to derive an embedding from a target voice. The embedding is low-dimensional and is a meaningful representation of the reference voice. Furthermore, The synthesizer, conditioned on the embedding, produces a spectrogram from the text and the target voice. The synthesizer features Tacotron, which is a recurrent sequence-to-sequence model, containing an encoder-decoder structure, that predicts and produces a mel-spectrogram. The vocoder utilises the spectrogram, generated by the synthesizer model, to deduce an audio waveform. The procedure of the speech module starts with the speaker encoder receiving a short reference voice of the target speaker. Afterwards, the encoder produces an embedding that conditions the synthesizer. The synthesizer is then fed a text to generate a spectrogram. In the synthesizer, the embedding is concatenated to each frame that is produced by the Tacotron encoder. Then the attention mechanism utilises the encoder output frames to synthesize the decoder input frames [5]. Every decoder input is concatenated with the previous decoder frame output, leading to the model being autoregressive. The whole sequence of frames goes through a residual post-net, eventually becoming the mel-spectrogram. The vocoder takes the spectrogram to produce a user-specified waveform similar to the short reference voice used as the input for the encoder.

2.1.2.2 Face Generation Module

Figure 4: Our proposed neural network architecture for speech-to-face landmarks module

The proposed neural network in Figure 4 of the landmark generation module followed the paper in [2]. It takes a short frame of the raw waveform (speech) as input and produces the Active Shape Models (ASMs) weights corresponding to the frame [3]. ASMs are deformable statistical models of the shape of objects that represent the variation in the training set by a set of coefficients. The coefficients and weight are obtained by Principal Components Analysis (PCA) [4]. The input will be processed by four 1D convolutional layers. The numbers of filters increase as the time dimensions diminish. Strides are used for each convolutional layer and followed by the Leaky-ReLU activation function with a slope of 0.3 and dropout layers with a dropout rate of 0.2. One fully connected layer follows the stack of convolutional layers that outputs the ASM’s weight.

Using ASMs, we can obtain the generalized face landmarks video by s = sμ + wS , while sμ is the mean shape vector, w is the vector of weight generated by the neural network, and S is a matrix that contains eigenvectors. The parameters of the neural network will be further modified if needed to achieve the best result.

Figure 5: Our proposed architecture for face landmarks-to-video module

Our person-specific realistic head model mapping module shown in Figure 5 follows a similar architecture as proposed in [5]. The proposed architecture consists of a motion module and a generation module. The motion module predicts the dense motion field from frames of the face landmarks to the source frame of still image or video. The predicted dense motion field is used to

align the features map from the source frame. Inside the motion module, there is a key point detector that extracts the sparse key points and local affine transformation with respect to an abstract reference frame, followed by a dense motion module that generates optical flow from face landmarks to source frame and occlusion map. The generation module takes the source image and the output from the dense motion module to generate the realistic talking head video. 2.2 Implementation 2.2.1 Dataset Collection

Figure 6a: The left side shows a video available on YouTube which shows a portrait view of Barack Obama which contains his face movements and speech. (Script included in Github)

Figure 6b: A glimpse of transcript for Figure 6a extracted using the YouTubeTranscriptApi. (Script included in Github) The proposed architecture for speech generation module requires training the encoder, the synthesizer, and the vocoder on the same set of dataset. The encoder requires a large number of speakers whereas transcripts are required for the synthesizer. The specific requirements on the datasets for each of the components narrowed our results on open-source datasets. Since small datasets are likely to lead to poor generalisation and reduce the performance of zero-shot, our team explored alternative ways to extract data which led us to YouTube. YouTube, being one of the largest online video sharing platforms, had videos that met the needs of the architecture. Therefore, our team handpicked several videos similar to Figure 6a as our data. We extracted the audio from the YouTube videos, as shown in Figure 6c, in order to use them for the speech generation module. Moreover, since the synthesizer requires a transcript for the audio in the video, the team used YouTubeTranscriptApi to generate a list of transcripts as shown in Figure 6b which satisfied the requirements of our dataset.

Figure 6c: An audio file and video file is extracted from the YouTube video shown in Figure 2a to train the architecture. Moreover, our face model generation module requires us to generate a face model given a sample of speech of the user. By extracting the audio file from the raw Youtube video, we have an audio-video dataset which can be used as our training data for the face model generation module. This process of using YouTube videos allows an easier collection of data that can be used by both of our modules. 2.2.2 Real-Time Voice Cloning To enhance the realistic attribute of our talking head model, voice recording and speech synthesis will be implemented. For our application to receive data in audio format, users can either load a pre-recorded audio or record their voice in real-time. The audio data gathered will be used to further train a pre-trained model, hence improving the coherency and similarity of the synthesised voice. With limited open-source projects related to direct voice cloning, we have opted to utilise the most prominent repository, Real-Time Voice Cloning [11]. The output of this project aligned with the goal of our Final Year Project, where we desire reducing the amount of data required from the user, whilst maintaining the quality of the synthesized audio. 2.2.2.1 Pretrained Models The open-source library provides three pre-trained models, the synthesizer, the speaker encoder and the vocoder. The encoder is trained to 1,560,000 steps with a batch size of 64, on 2500 hours of English speech extracted from audiobooks and celebrity interviews from YouTube. For both the

synthesizer and vocoder model, they are each trained on 460 hours of English speech derived from audiobooks. However, the synthesizer is trained to 278,000 steps and the vocoder is trained to 428,000 steps. 2.2.2.2 Preprocessing training audio data From the data gathered from 2.2.1, we divided the audio file into separate files, each with a duration of around 7 seconds. Afterwards, we created a transcript for each of those files.

Figure 7: separate audio and transcript files sorted by number and type.

Furthermore, we arranged the files into a specific format, identical to Figure 7, so that it can be read easily when training the dataset. 2.2.2.3 Speech Model Modification and Improvement We decided to finetune a single-speaker model for each user by generating our training data for each voice loaded. Instead of training from scratch, we decided to train on top of the pre-trained model that is already trained to 278,000 steps. This proved to be more beneficial as the model has already generalized with numerous voices and accents on English speech. From our observation, around 10 utterances (100 seconds of audio) and an additional 1000 steps are the minimum requirements to produce satisfying results. Moreover, we have learned that fine-tuning the synthesizer model has improved both the coherence and similarity with the target voice and replicates the target’s accent as well. On the other hand, fine-tuning the vocoder has proven to not be as impactful compared to training the synthesizer. Therefore, we have decided to focus on solely fine-tuning the synthesizer model. This will lead to generating a more accurate and similar voice for the user in real-time. Initially, new voices foreign to our model would have to be trained, but afterward, users can clone in real-time for future use.

We have resolved the potential problem of gap occurrence by using voice activation detection, webrtc Voice Activity Detector, to trim excess silence in our training data.

Figure 8: Spectrogram, with reduced gaps, produced after fine-tuning synthesizer.

To achieve the results in Figure 8, we preprocessed LibriSpeech, by making the minimum seconds of silence accepted from 0.4 to 0.05 seconds. This leads to the breaking up of all the utterances that contain long pauses since 0.4 seconds is too long. After trimming the silence from each utterance we compare the before and after. If the duration does not match, then silence is detected and thus the old utterance, before being trimmed, is deleted. Overall the project was able to automatically trim any long pauses in the training data that affected the spectrogram and reduce the gaps. 2.2.3 Realistic Head Model Generation 2.2.3.1 Pre-trained Models We obtained the open-source library of the face landmarks generation model and the realistic head model generation model, which was implemented according to the paper [2][5]. Two networks were trained and provided the pre-trained checkpoints. The face landmarks neural network was implemented in PyTorch and was trained for 100 epochs. The mini-batch size and learning rate were set to 128 and 10e−4, respectively. For the realistic head model generation, the neural network was trained with Adam optimizer with learning rate 2e − 4 and batch size 20 and employed learning decay by dropping the learning rate at 50000 and 75000 iterations. Both models are able to produce 25 frames per second video. We will train on top of the pre-trained checkpoint with the dataset mentioned. 2.2.3.2 Preprocessing speech signal data Speech signals processing often encounters the problem of containing unwanted noise that worsens the performance of an automatic speech processing system. For the speech to face landmarks module, it is necessary to apply a signal filter to extract more important parts of data before we feed it into the neural network, which increases the performance of the face landmarks module.

Figure 9: trimmed original spectrogram of input speech from previous text-to-speech module (left), and spectrogram of filtered speech (right) After applying the filter with a cut-off frequency of 4000 Hz, we enhanced the voice signal below the cut-off frequency, reducing the effectiveness of the high-frequency signal. We input both unfiltered and filtered voice of the spectrogram shown in Figure 9 into the face landmark generation network separately and found that there is a small improvement in the landmark video, which the eyelid had less shaking inside the filtered video with the speech speaking at flat intonation, thus reducing unusual shaking and producing more realistic talking face movement. Another preprocessing concern about the input image of the face landmarks to the video module. As the model architecture require the resolution of an input image to be 256*256, which is usually smaller than other regular image resolution. We need to resize and scale down the image. However, it comes with a potential problem in that it may change the ratio of the image and make it look unrealistic. It is necessary to keep the header image with the same ratio to generate a realistic talking head video. Considering the potential problem, add zero paddings to the corresponding border (“top and bottom” or “left and right”) before resizing the image can maintain the ratio of the image and fulfill the requirement of the input image. Considering the output from the neural network model of landmarks-to-video module, the mouth and eye movement will cause the background to “vibrate” as the background does not change with the face movement, which looks unrealistic. Removing the background and changing it to white color can minimize the vibration of the background. This can be an optional improvement and choice for the user to choose whether to keep the background. Figure 10 shows an example.

Figure 10: with background (left) and without background (right)

2.2.3.3 Model Modification and improvement As to obtain a better result, we have done some experiments on the face landmarks generation, through limiting face movement. The 2D landmarks model outputs a set of coordinates that represent the current position of features. It utilizes the matplotlib to plot the face landmarks for each frame and combine them into an output video. We implemented movement limitation which reduces the change in coordinate according to the previous coordination, stabilizing the outline and reducing unnecessary features for example nose and eyebrows. We have conducted the above changes and experiments with different combinations.

2.2.4 Mobile Application 2.2.4.1 Frontend For the User Interface, we focused on displaying a friendly user interface for the user and easy upload and retrieval of data from the server. Our application prompts the user for an image or a video of them with their facial features clearly shown by either selecting a photo from their albums or taking their photo through the application directly. We also require the user to input an inquiry or request in text format. The frontend will gather both data and post it via Fetch to the server. Currently, we have developed an initial prototype for all the necessary views for our mobile application using Adobe XD. Figure 11 shows some of the major frames of our initial prototype for the application, created using Adobe XD, which outlines the major features of the application.

Figure 11: Application prototype that shows the main tasks involved in the frontend. The camera mode (left) will prompt users initially to record or select a source video. The home (right) will allow users to input text and will display the generated visual video.

To retrieve the results from the server, we will use the Fetch API. Fetch provides us with an interface for fetching data. The React Native community provides us with a library called react-native-camera that presents an easy interface to interact with a phone’s camera. To upload the recorded video, the data will be put into a FormData() object and will be sent via post to the server through Fetch. When retrieving the end results of the Text-to-Video, we again use Fetch to receive the video from the server and present the video on the screen. 2.2.4.2 Backend The backend for the application consists of three major tasks including communicating with the frontend, and deploying our realistic head model generation system. In order to facilitate the backend communication with the mobile application, the backend will be carried out in a virtual machine given by the Department of Computer Science and Engineering (cssystem). In the virtual machine, we will use a web framework named Flask to write our server-side script with Docker Compose which provides a convenient way to configure multiple Docker containers and services [6][7]. In addition to this, NGINX will act as our HTTP webserver to host our model

with the help of Flask [8]. In order to make sure the web application is able to scale up, recover from a crash, and handle HTTP requests, Gunicorn which is a Web Server Gateway Interface (WSGI) is used with Supervisor [9][10]. The server-side Flask implementation will handle our work on realistic head generation and return the result back to the client that requested through the Flask API call. Figure 12 shows an overview of the main components involved in the backend.

Figure 12: Overview of the backend. The client (mobile application) requests data (realistic head model) through the Nginx web server. Gunicorn handles proxies between Nginx and the web framework Flask. The pre-trained model will be handled in the Flask. A supervisor will monitor and control Gunicorn by restarting the process in case additional features are implemented or in case of unexpected errors.

3. Discussion 3.1 Results 3.1.1 Real-time Voice Cloning We conducted trials, testing different numbers of steps and utterance to produce a coherent voice similar to the reference speech inputted. Based on training data size As mentioned above, loading an utterance that the model has not generalised with, frequently leads to huge gaps in the spectrogram.

Figure 13: (a) Synthesizer pre-trained model, trained to 278,000 steps on LibriSpeech. (b) Fine-tuned synthesizer model, trained to 279,000 steps on 5 utterances of a target speech. (c) Fine-tuned synthesizer model, trained to 279,000 steps on 10 utterances of a target speech. (d) Fine-tuned synthesizer model, trained to 279,000 steps on 20 utterances of a target speech. (e) Fine-tuned synthesizer model, trained to 279,000 steps on 40 utterances of a target speech.

To indicate the minimum utterance needed to synthesize a coherent speech, we conducted a trial across different numbers of utterances. To make our experiment fair, we used the same target voice and trained each to 279,000 steps, 1000 steps on top of the pre-trained model. From the first trial, we identified that 5 or more utterance leads to reducing the gaps in the spectrogram significantly. However, we were dissatisfied with the quality of the voice from Figure 13 (a), as it sounded unrealistic. This may be due to the spectrogram not containing any small pauses in the first few seconds, leading to the speech being too fast and incoherent. Training on only 10 utterances, Figure 13 (b), we found that both the quality and similarity to the target voice drastically improved. After training on 20 utterances, Figure 13 (c), there was a slight improvement in the quality of the voice. When training on 40 utterances, Figure 13 (d), we found the same pattern, where the coherency and similarity to the target voice improved moderately. To conclude our test, we were already pleased with the quality of the synthesised speech from only using 10 utterances (approximately 1 minute of audio). Therefore, for the application of our

project, future users will have to initially input 10 sentences minimum of the target voice to receive a realistic cloned speech. Based on the number of training steps To identify the minimum number of steps needed to produce satisfying results, we conducted a trial across different numbers of steps. To make our experiment fair, we used the same target voice and trained on 10 utterances for each trial.

Figure 14: (a)Fine-tuned Synthesizer pre-trained model, trained to 279,000 steps on 10 utterances. (b) Fine-tuned synthesizer model, trained to 279,500 steps on 10 utterances of a target speech. (c) Fine-tuned synthesizer model, trained to 280,000 steps on 10 utterances of a target speech. (d) Fine-tuned synthesizer model, trained to 280,500 steps on 10 utterances of a target speech. (e) Fine-tuned synthesizer model, trained to 281,000 steps on 10 utterances of a target speech. (f) Fine-tuned synthesizer model, trained to 281,500 steps on 10 utterances of a target speech.

In conclusion, as the steps increase, the coherency and similarity to the target voice improved. When comparing Figure 14 (a) and Figure 14 (f), there is a noticeable difference, the slight gaps.

The slight gaps increase authenticity as it makes the audio waveform sound less robotic due to replicating natural pauses in human speech. However, when showcasing these audio results to a group of people, they were already satisfied with 279,000 steps. Therefore, we decided to use 279,000 steps as the minimum requirement for training the dataset. 3.1.2 Realistic Head Model Generation With the changes made on the face landmarks generation model, we found there is a significant improvement in the talking-head video production with the video nearly not having exaggerated motion on the face while keeping the essential facial expression on the speech of different intonation. Without the modification, the inference video had the head moving backward as shown in Figure 15 below, which produced inconsistent movement in the video, degrading the overall quality of the result.

Figure 15: The frames that saying the word “course” with high intonation generated with original and modified features plotting face landmarks video 3.1.3 Mobile application Demo The Figure 16 below shows the execution of our application in the frontend. Users can press the “generate” button to generate the head model video. Then the user can select a trained speech model from the list and take a picture in portrait view. The picture will be uploaded to the backend to set up the generation modules. After sending the picture, users can type in text that can be spoken in the talking head model video. The text will upload to the backend for processing and producing the video. Last, the application receives the talking-head video from the backend and displays it on the screen. Users can repeat many times in the input text steps to produce more videos. For the speech model training, users can first press the “record” button to create a new model. It will require the user to enter the model name. Then, it will open the record audio interface and the users are required to read 10 sentences for training dataset collection.

3.2 Evaluation

Figure 16: System Flow of our application Mynah

The team evaluated the system through the means of mean opinion score to provide concrete validity of our results. A group of volunteer evaluators, around 20 people, have been asked to give ratings from 0 to 5 for the accuracy of the generated video on aspects of speech and video generation. Among 20 users, a mean score of 4.85 was given for the generated speech whereas a mean score of 4.05 was given for the generated video.

Furthermore we also surveyed members of the target groups, the verbally impaired population, by asking about the usefulness of our system, for which the responses were mostly positive.

We were fortunate to have the opportunity to interview Jonathon, who is diagnosed with ADHD and Autism. With the difficulties he encounters regarding communication, Jonathon is one of the potential target consumers of our project. Therefore we asked him questions about the challenges he faces when interacting with society and his opinion on the output of our application. Jonathan concluded that our application has the potential to provide relief for him and people who face similar difficulties, as it can help relay their messages to others with ease, especially in important scenarios such as job interviews.

4. References

[1] S. M. Pascua, P. L. C. Espina, R. P. E. Talag, L. N. Villegas, and L. Aquino de Guzman, “Words in vision : A Filipino sign language THESAURUS management system Using ren-py,” 01-Jan-1970. [Online]. Available: http://library.ifla.org/1732/. [Accessed: 04-Mar-2021].

[2] S. E. Eskimez, R. K. Maddox, C. Xu and Z. Duan, 'Noise-Resilient Training Method for Face Landmark Generation From Speech,' in IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 28, pp. 27-38, 2020. doi: 10.1109/TASLP.2019.2947741. [Accessed 7 February 2021]

[3] T. F. Cootes, C. J. Taylor, D. H. Cooper, and J. Graham, “Active shape model-Their training and application,' Comput, Vision Image Under-standing, vol. 61, no. 1,pp. 38-59,1995 [Accessed 8 February 2020]

[4] I. T. Jolliffe, J. Cadima(2016), 'Principal component analysis: a review and recent developments.” Phil. Trans. R. Soc. A, 374:20150202. [online] Available at: https://royalsocietypublishing.org/doi/10.1098/rsta.2015.0202 .[Accessed 11 Sep. 2020]

[5] A. Siarohin, S. Lathuiliere, S. Tulyakov, E. Ricci, N. Sebe(2019). First Order Motion Model for Image Animation. [online] Available at: https://papers.nips.cc/paper/8935-first-order-motion-model-for-image-animation.pdf [Accessed 1 Sep. 2020]

[6] “Welcome to Flask¶.” [Online]. Available: https://flask.palletsprojects.com/en/1.1.x/. [Accessed: 17-Sep-2020].

[7] “Overview of Docker Compose,” Docker Documentation, 16-Sep-2020. [Online]. Available: https://docs.docker.com/compose/. [Accessed: 17-Sep-2020].

[8] “High Performance Load Balancer, Web Server, & Reverse Proxy,” NGINX, 13-Sep-2020. [Online]. Available: https://www.nginx.com/. [Accessed: 17-Sep-2020].

[9] “Gunicorn 'Green Unicorn' is a Python WSGI HTTP Server for UNIX. It's a pre-fork worker model. The Gunicorn server is broadly compatible with various web frameworks, simply implemented, light on server resources, and fairly speedy.,” Gunicorn. [Online]. Available: https://gunicorn.org/. [Accessed: 17-Sep-2020].

[10] “Hello Linux: Gunicorn & Supervisor // Blog // Coding for Entrepreneurs,” Coding for Entrepreneurs. [Online]. Available: https://www.codingforentrepreneurs.com/blog/hello-linux-setup-gunicorn-and-supervisor. [Accessed: 17-Sep-2020].

[11] C. Jemin, (2018). Real-time Voice Cloning [online] Available at: https://puu.sh/DHgBg.pdf [Accessed 15 Jan. 2021]