The understanding of emotions from voice by a human brain are normal instincts of human beings, but automating the process of emotion recognition from speech without referring any language or linguistic information remains an uphill grind. In the research work presented based on the input speech, I am trying to predict one of the six types of emotions (sad, neutral, happy, fear, angry, disgust). The diagram given below explain how emotion recognition from speech works. The audio features are extracted from input speech, then those features are passed to the emotion recognition model which predicts one of the six emotions for the given input speech.
Working Of Emotion Recgnition From Speech
Motivation
Most of the smart devices or voice assistants or robots present in the world are not smart enough to understand the emotions. They are just like command and follow devices they have no emotional intelligence. When people are talking to each other based on the voice they understand situation and react to it, for instance if someone is angry then other person will try to clam him by conveying in soft tone, these kind of harmonic changes are not possible with smart devices or voice assistants as they lack emtional intelligence. So adding emotions and making devices understand emotions will take them one step further to human like intelligence.
Application
There are tonnes of applicates based on one can imagine. Few applications based on my thinking are human computer interaction using voice, home automation, anger/stress management by decoding emotions from voice, emotion recognition can help in detecting fear and cops can used this system to check if dialer is feared by some one or its just a normal call to register a complain, Marketing companies can use emotions to sell products based on user mood, autonomous vehicles can detect user emotion and adjust the speed of vehicles, It can help in solving psychological or depression problems. These are few applications according to me but there can be many more as voice based systems are increasing, even voice bsed chatting is common on social media platforms like clubhouse, discord, twitch, and others.
Libraries and coding language used for the project
languages
programming style
programming language
Dataset description
I have used four datasets and all four datasets are freely available to downloaded from kaggle website. So I have downloaded the data, extracted and stored in my google drive.
Ryerson Audio Visual Database of Emotional Speech and Song (Ravdess) dataset description:
Dataset link to download: "https://www.kaggle.com/uwrfkaggler/ravdess-emotional-speech-audio"
Dataset stored on google drive at path: "/content/drive/MyDrive/Audiofiles/audio_speech_actors_01-24/"
Dataset contains sub folders and file names as example in numbers format 03-01-01-01-01-01-01.wav.
Actor (01 to 24. Odd numbered actors are male, even numbered actors are female).
So based on the number there is a identifier for each number and its meaning are as follows:
Therefore file 03-01-01-01-01-01-01.wav can be deduced as 03=audio-only, 01=speech, 01=neutral, 01=normal, 01=statement kids and 01=1st repetition.
Crowd sourced Emotional Mutimodal Actors Dataset (CREMA-D) dataset description:
Dataset link to download: "https://www.kaggle.com/ejlok1/cremad"
Dataset stored on google drive at path: "/content/drive/MyDrive/Audiofiles/AudioWAV/"
The format of files is 1001_DFA_ANG_XX.wav, where ANG stands for angry emotion.
Similarly different emotion mappings are as follows:
{'SAD':'sad','ANG':'angry','DIS':'disgust','FEA':'fear','HAP':'happy','NEU':'neutral'}
Toronto emotional speech set (Tess) dataset description:
Dataset link to download: "https://www.kaggle.com/ejlok1/toronto-emotional-speech-set-tess"
Dataset stored on google drive at path: "/content/drive/MyDrive/Audiofiles/TESS Toronto emotional speech set data/"
There are folders in format OAF_angry, OAF_neural, OAF_disgust, YAF_sad and so on, where name after the underscore of the folder name contains the emotion information, so the name after the underscore of the folder name is taken and files residing insider the folders are labeled accordingly.
Surrey Audio Visual Expressed Emotion (Savee) dataset description:
Dataset link to download: "https://www.kaggle.com/ejlok1/surrey-audiovisual-expressed-emotion-savee"
Dataset stored on google drive at path: "/content/drive/MyDrive/Audiofiles/ALL/"
The files are in a format DC_a01.wav where a single character contains the emotion information , for example character 'a' after underscore in the file name "DC_a01.wav" means emotion is angry.
Similarly different emotion mappings are as follows:
{'a':'anger','d':'disgust','f':'fear','h':'happiness','n':'neutral','sa':'sadness','su':'surprise'}
Universal decorator fucntion to calculate total time
def calc_time(func):
def inner(*args, **kwargs):
st = time.time()
result = func(*args,**kwargs)
end = time.time()-st
print("Total time required: {:.3f} ms".format(end * 1000))
return result
return inner
Description of important functions present in code (Model design and evaluation):
There are many functions in the program as functional programming style is used. Here I am going to describe a few important functions which call other functions inside the functions and generate files and results. Detailed description of each function and its use can be found in the code file.
Audio_features_extract() this function is used to extract audio features and generates a csv file at path "/content/drive/MyDrive/Audiofiles
/Audio_features_All_pr.csv" which contains audio features and their respective label information.
Below given image shows snapshot of the csv file, the file has a total of 33954 rows × 179 columns.
csv file snapshot
The csv file is loaded using pandas library, additional_preprocess() function carries out Exploratory Data Analysis and drop emotions with limited samples to
avoid missclassifications and then dataset is divided into train, test and validation set.
Below image gives the detailed description of the whole process.
Explorator Data Analysis and data preprocessing
Deep learning model for speech recognition is trained using the training data and at every epoch or checkpoint validation accuracy is calucated. The epoch or
checkpoint which gives highest validation accuracy, the best model is saved for that epoch or checkpoint at path "/content/drive/ MyDrive/Audiofiles/
emotionrecognition.hdf5", the model giving highest validation accuracy is only saved.
model training snap shot
Description of testing model in real time:
Once the model is build and training is completed the emotion recognition model can be loaded from the path "/content/drive/MyDrive/Audiofiles/emotion-recognition .hdf5" and can be tested for the given input speech in real time.
The data for real time model testing is recorded using the microphone.
Then features are extracted from speech and passed to emotion recognition model which predicts one of the six emotions.
Below figure shows the audio waveform and output of the emotion recognition model.
realtimeresult
Results
Below figure shows the training, testing and validation accuracy achieved by the emotion recognition model.
accuracy
Below figure shows the classification report and it can be seen in the report that for all the classes the value is greater than 0.5 which means the model can predict the emotions accuratly to some extent. If the value is equal to 1.0 for all clases then it means model can predict accurrately always given the input speech. But its diffcult to achieve real time prediction and 100% accuracy on real time envoiurment as there is noise and many other factors which can affect the output. Given a challenge it can overcomed by training with big set of data in different languages to develop a universal model.
classification
The 0,1,2,3,4,5 in classification report resembles to different emotions which can be decoded from below image.
emotionsmapping
Below figure shows output of confusion matrix.
confumatrix
Below figure shows the training loss and accuracy curves, despite the model giving the training accuracy of 100%, validation and testing accuracy is near to 75%-76%, my model gives the highest accuracy when compared to the authors who previously carried out the research work in this area.
curves
Additionally to check wheather the model can work for all types of voices and on unlabeled data a test was carried out using combination of different voices and unlabled data. Below figure shows the results.
unlabeltest
Below figures shows comparison of my model with other authors who worked previously in this area of emotion recognition from speech.
comparison
Installation
To download and run my google colab file 1130532_ResearchMethodology_Project_Final.ipynb following changes need to be made:
Frist and foremost make sure all neccessary libraries mentioned above are installed.
To install any library in the computer machine just use command pip install library name.
Extract the downloaded data from the above given links
Once the data is extracted just use my code and pass the proper path information to the functions.
These paths are datapaths, csv file path and paths where reults are stored.
Correct path information needs to be given in the functions ravdess_data(), crema_data(), tess_data(), saveee_data(), fetch_data(), Audio_features_extract(), audio_features_final(), emotion_recognition_model(), test_realtime(), evaluate_model(), unknown_audio() and diff_lang_test()
Below given are snnipets of code where proper path information needs to be given for the above given functions.
So once the path information is given correctly its time to run the functions, run all the fuctions in the same sequence given in my colab file 1130532_ResearchMethodology_Project_Final.ipynb.
If one dosent want to train the model just test the model then they can use the model file "emotion-recognition.hdf5", change the paths in test_realtime()
function and they can test the model.
If you want to develop or implement or setupt the whole code then as mentioned give proper paths and run all the functions its done.
Check out my colab file 1130532_ResearchMethodology_Project_Final.ipynb to see the time required by the individual process to complete.
The main() function does all the work of training the model and evaluating the model. Once the main function completes running the model is file is generated and can used for real time testing.
This is all about installation, building the model and feature extraction are one time process, once completed model is deployed in real time enviourment for testing and using the model for recognizing emotions from speech.
usage
As mentioned in the installation process, once libraries, datasets are downloaded, proper path information is given functions should be run in a sequence as mentioned in the colab file 1130532_ResearchMethodology_Project_Final.ipynb.
Following are the functions that required to run and the sequence is same as mentioned below and in the colab file 1130532_ResearchMethodology_Project_Final.ipynb
Remeber that every function requires amount of time to complete the process so.
Following is the list of sequence of functions which are required to run after running the import libraries code cell section:
Universal python decorator function to calculate total time.
Below given functions are for feature extraction, run this functions only once as it requires time to extract features form auido. Also features extraction is a one time process. Once features are extracted we can carry out further processing and train the emotion recognition model.
main() function calls the functions in a sequence and after the execution of the main() function the deepl learning model for emotion recognition is ready.
@calc_time
def main():
#get train,test data and labels
x_train, x_test, y_train, y_test, x_val, y_val, encoder = audio_features_final()
#call the emotion recognition model
emotion_recognition_model(x_train,y_train,x_val,y_val)
#evaluate the model performance
evaluate_model(x_train, x_test, y_train, y_test, x_val, y_val)
if __name__:main()
Once the model is trained and model file is generated one can use the below fucntions to test the model in real time enviourment.
If some one wants to used my trained model file directly then no need to run the main() function just run the above given two function to test in the real time enviourment.
Also if using google colab make sure the function which are called inside the audio_features_final() and realtime_tested() are executed in advance as these two functions are dependent on them.
Make sure all functions are called properly as mentioned in my colab file 1130532_ResearchMethodology_Project_Final.ipynb
Additionally for my research work I carried on unkown sample data in different languages.
You will require to preprocess the data then you can use my get_features_recorded(audio,sr) function to get the audio features then pass the audio features to the
model to predict the outcome.
I have already downloaded the few audio samples for testing on different voices and data is available on my google drive link, please sendme mail to access the data I will give acccess to the google dr
For my custom data in different languages I have used below functions to test the emotion recognition models.
def unknown_audio()
def diff_lang_test()
'''
Whenever using the code make sure the function used inside the fnctions are called prior to executing the required function and all functions are executed in a proper sequence.
Video on Installation / Usage
Below given is the video on installation and usage of my project code in colab file. The video guides you by showing where to make path changes and how to install and run the code and test the code. The datasets link to download are already given above. I would request to read the installation and usage section and then watch this video so it will give a clear idea of the whole project and its working. click on the below youtube image to launch the video.
Support / Contact details
Given below are few of my social media accounts where anyone can contact me.
One can also contact me by email
For access to my google drive to see the setup of the whole project mail me on gmail id mentioned above access will be given to the selected people for some amount of time.
Road-map (future ideas)
The backgorund noise may cause errors when testing the model in real time enviourment and thus it can affect the output of the model. To avoid the noise audio segmentation needs to be performed, so I am planning to develop an audio segmentation model which can seprate user speech from background noise so emotions can be predicted accurately. Also I will be collecting audio in different formats extract features and train the model so a universal model can be developed. Once audio model is build it can be applied to video also by combining audio model of emotion recognition with facial model for emotion recognition, this can help in acheving more accurate output. Additionally three models can be combined that is textual, voice and facial based but it requires huge computation power and there is very limited study available on combining three models for emotion recogniton, beaucse a avoting mechanism or strategy needs to be developed for predicting the emotion from three models as there can be cases where each model can predict different emotions or two model predict same emotion and one predicts another emotion. Moreover I would like to build a audionet kind of embeddings similar to imagenet and word embeddings which will help other researchers working in this area to use pretrained audio embeddings.
How to contribute
One can contribute by extracting features from different auido files the code for extracting features can be used from my ipynb file, different dataset may reqire different data preprocessing so one also write a function for data preprocessing and send me both prerporcessing code and csv file, so I can integrate both data preprocessing function and csv file with my csv file Audio_features_All_pr.csv. Additionally I am planning to build three model audio segmentation model, facial emotion recognition model and textual model so one can contribute by writing the function for the same and integrate it. Send me a git merge request to integrate code or contact me so we can check the integrity of code and combine the code. One can also branch out create their own branch and then we can merge the branch. Additionaly one can also fork the repository.
I would like to thank @Ricardo for providing javascript code to inovke mircophone of user from google colab cell. As google colab dosent support audio recording using microphone so a javacript function needs to be written to inovke microphone and record auido. I would also like to thank @Fadi Badine my deep learning neural network model for emotion recognition is based on his model for automatic speech recognition.
References
[1] Francesc Alı́as, Joan Claudi Socoró and Xavier Sevillano, ”A Review of Physical and Perceptual Feature Extraction Techniques for Speech, Music and Environmental Sounds”, Appl. Sci. 2016.
[2] Kannan Venkataramanan and Haresh Rengaraj Rajamohan, ”Emotion Recognition from Speech”, arXiv:1912.10458v1 [cs.SD] 22 Dec 2019.
[3] Haiyang Xu, Hui Zhang, Kun Han, Yun Wang, Yiping Peng and Xian-gang Li, ”Learning Alignment for Multimodal Emotion Recognition from Speech”, arXiv:1909.05645v2 [cs.CL] 3 Apr 2020.
[4] Aharon Satt, Shai Rozenberg and Ron Hoory, ”Efficient Emotion Recognition from Speech Using Deep Learning on Spectrograms”, INTERSPEECH 2017, Stockholm, Sweden, August 20–24, 2017.
[5] Jia Rong, Gang Li and Yi Ping Phoebe Chen, ”Acoustic feature selection for automatic emotion recognition from speech”, Information Processing and Management 45 (2009) 315–328.
[6] K. Sreenivasa Rao, Tummala Pavan Kumar, Kusam Anusha, Bathina Leela, Ingilela Bhavana and Singavarapu V.S.K. Gowtham, ”Emotion Recognition from Speech”, (IJCSIT) International Journal of Computer Science and Information Technologies, Vol. 3 (2) , 2012,3603-3607.
[7] Vladimir Chernykh and Pavel Prikhodko, ”Emotion Recognition From Speech With Recurrent Neural Networks”, arXiv:1701.08071v2 [cs.CL] 5 Jul 2018.
[8] Sabur Ajibola Alim and Nahrul Khair Alang Rashid, ”Some Commonly Speech Feature Feature Extraction Algorithms”. Published: December 12 2018, DOI: 10.5772/intechopen.80419.
[9] Oh Wook Kwon, Kwokleung Chan, Jiucang Hao and Te Won Lee, ”Emotion Recognition by Speech Signals”, GENEVA, EUROSPEECH 2003.
[10] K.V.Krishna Kishore and P.Krishna Satish, ”Emotion Recognition in Speech Using MFCC and Wavelet Features”, IEEE International Advance Computing Conference (IACC), 2013.
[11] Panagiotis Tzirakis, Jiehao Zhang and Björn W. Schuller, ”END-TO-END SPEECH EMOTION RECOGNITION USING DEEP NEURAL NETWORKS”, IEEE International Advance Computing Conference (IACC), 2018.
Emotion recognition model is finished and its ready and can be used in real time.
The 1130532_ResearchMethodology_Project_Final.ipynb file can be downloaded and used by providing neccesary path changes as mentioned in installation and usage sections.
I am looking forward to develop other models mentioned in road-map (future ideas) and integrate all those models with my current emotion recognition model.
