- This repository handles building and training Speech Emotion Recognition System.
- The basic idea behind this tool is to build and train/test a suited machine learning ( as well as deep learning ) algorithm that could recognize and detects human emotions from speech.
- This is useful for many industry fields such as making product recommendations, affective computing, etc.
- Check this tutorial for more information.
- Python 3.6
- librosa==0.6.3
- numpy
- pandas
- soundfile==0.9.0
- wave
- sklearn
- tqdm==4.28.1
- matplotlib==2.2.3
- pyaudio==0.2.11
- ffmpeg (optional): used if you want to add more sample audio by converting to 16000Hz sample rate and mono channel which is provided in
convert_wavs.py
Install these libraries by the following command:
pip3 install -r requirements.txt
This repository used 4 datasets (including this repo's custom dataset) which are downloaded and formatted already in data
folder:
- RAVDESS : The Ryson Audio-Visual Database of Emotional Speech and Song that contains 24 actors (12 male, 12 female), vocalizing two lexically-matched statements in a neutral North American accent.
- TESS : Toronto Emotional Speech Set that was modeled on the Northwestern University Auditory Test No. 6 (NU-6; Tillman & Carhart, 1966). A set of 200 target words were spoken in the carrier phrase "Say the word _____' by two actresses (aged 26 and 64 years).
- EMO-DB : As a part of the DFG funded research project SE462/3-1 in 1997 and 1999 we recorded a database of emotional utterances spoken by actors. The recordings took place in the anechoic chamber of the Technical University Berlin, department of Technical Acoustics. Director of the project was Prof. Dr. W. Sendlmeier, Technical University of Berlin, Institute of Speech and Communication, department of communication science. Members of the project were mainly Felix Burkhardt, Miriam Kienast, Astrid Paeschke and Benjamin Weiss.
- Custom : Some unbalanced noisy dataset that is located in
data/train-custom
for training anddata/test-custom
for testing in which you can add/remove recording samples easily by converting the raw audio to 16000 sample rate, mono channel (this is provided increate_wavs.py
script inconvert_audio(audio_path)
method which requires ffmpeg to be installed and in PATH) and adding the emotion to the end of audio file name separated with '_' (e.g "20190616_125714_happy.wav" will be parsed automatically as happy)
There are 9 emotions available: "neutral", "calm", "happy" "sad", "angry", "fear", "disgust", "ps" (pleasant surprise) and "boredom".
Feature extraction is the main part of the speech emotion recognition system. It is basically accomplished by changing the speech waveform to a form of parametric representation at a relatively lesser data rate.
In this repository, we have used the most used features that are available in librosa library including:
- MFCC
- Chromagram
- MEL Spectrogram Frequency (mel)
- Contrast
- Tonnetz (tonal centroid features)
Grid search results are already provided in grid
folder, but if you want to tune various grid search parameters in parameters.py
, you can run the script grid_search.py
by:
python grid_search.py
This may take several hours to complete execution, once it is finished, best estimators are stored and pickled in grid
folder.
The way to build and train a model for classifying 3 emotions is as shown below:
from emotion_recognition_using_speech.emotion_recognition import EmotionRecognizer
from sklearn.svm import SVC
# init a model, let's use SVC
my_model = SVC()
# pass my model to EmotionRecognizer instance
# and balance the dataset
rec = EmotionRecognizer(model=my_model, emotions=['sad', 'neutral', 'happy'], balance=True, verbose=0)
# train the model
rec.train()
# check the test accuracy for that model
print("Test score:", rec.test_score())
# check the train accuracy for that model
print("Train score:", rec.train_score())
Output:
Test score: 0.8148148148148148
Train score: 1.0
In order to determine the best model, you can by:
# loads the best estimators from `grid` folder that was searched by GridSearchCV in `grid_search.py`,
# and set the model to the best in terms of test score, and then train it
rec.determine_best_model(train=True)
# get the determined sklearn model name
print(rec.rmn.__class__.__name__, "is the best")
# get the test accuracy score for the best estimator
print("Test score:", rec.test_score())
Output:
MLPClassifier is the best
Test Score: 0.8958333333333334
Just pass an audio path to the rec.predict()
method as shown below:
# this is a neutral speech from emo-db
print("Prediction:", rec.predict("data/emodb/wav/15a04Nc.wav"))
# this is a sad speech from TESS
print("Prediction:", rec.predict("data/tess_ravdess/validation/Actor_25/25_01_01_01_mob_sad.wav"))
Output:
Prediction: neutral
Prediction: sad
from deep_emotion_recognition import DeepEmotionRecognizer
# initialize instance
# inherited from emotion_recognition.EmotionRecognizer
# default parameters (LSTM: 128x2, Dense:128x2)
deeprec = DeepEmotionRecognizer(emotions=['angry', 'sad', 'neutral', 'ps', 'happy'], n_rnn_layers=2, n_dense_layers=2, rnn_units=128, dense_units=128)
# train the model
deeprec.train()
# get the accuracy
print(deeprec.test_score())
# predict angry audio sample
prediction = deeprec.predict('data/validation/Actor_10/03-02-05-02-02-02-10_angry.wav')
print(f"Prediction: {prediction}")
Output:
0.7948717948717948
Prediction: angry
Predicting probabilities is also possible (for classification ofc):
print(deeprec.predict_proba("data/emodb/wav/16a01Wb.wav"))
Output:
{'angry': 0.8502438, 'sad': 1.15252915e-05, 'neutral': 8.986728e-05, 'ps': 0.14671412, 'happy': 0.0029406736}
print(deeprec.confusion_matrix(percentage=True, labeled=True))
Output:
predicted_angry predicted_sad predicted_neutral predicted_ps predicted_happy
true_angry 92.307693 0.000000 1.282051 2.564103 3.846154
true_sad 12.820514 67.948715 3.846154 6.410257 8.974360
true_neutral 3.846154 8.974360 82.051285 2.564103 2.564103
true_ps 2.564103 0.000000 1.282051 83.333328 12.820514
true_happy 20.512821 2.564103 2.564103 2.564103 71.794876
This repository can be used to build machine learning classifiers as well as regressors for the case of 3 emotions {'sad': 0, 'neutral': 1, 'happy': 2} and the case of 5 emotions {'angry': 1, 'sad': 2, 'neutral': 3, 'ps': 4, 'happy': 5}
- SVC
- RandomForestClassifier
- GradientBoostingClassifier
- KNeighborsClassifier
- MLPClassifier
- BaggingClassifier
- Recurrent Neural Networks (Keras)
- SVR
- RandomForestRegressor
- GradientBoostingRegressor
- KNeighborsRegressor
- MLPRegressor
- BaggingRegressor
- Recurrent Neural Networks (Keras)
You can test your own voice by executing the following command:
python test.py
Wait until "Please talk" prompt is appeared, then you can start talking, and the model will automatically detects your emotion when you stop (talking).
You can change emotions to predict, as well as models, type --help
for more information.
python emotion_recognition_using_speech/test.py --help
Output:
usage: test.py [-h] [-e EMOTIONS] [-m MODEL]
Testing emotion recognition system using your voice, please consider changing
the model and/or parameters as you wish.
optional arguments:
-h, --help show this help message and exit
-e EMOTIONS, --emotions EMOTIONS
Emotions to recognize separated by a comma ',',
available emotions are "neutral", "calm", "happy"
"sad", "angry", "fear", "disgust", "ps" (pleasant
surprise) and "boredom", default is
"sad,neutral,happy"
-m MODEL, --model MODEL
The model to use, 8 models available are: "SVC","AdaBo
ostClassifier","RandomForestClassifier","GradientBoost
ingClassifier","DecisionTreeClassifier","KNeighborsCla
ssifier","MLPClassifier","BaggingClassifier", default
is "BaggingClassifier"
This will only work if grid search is performed.
from emotion_recognition import plot_histograms
# plot histograms on different classifiers
plot_histograms(classifiers=True)
Output:
A Histogram shows different algorithms metric results on different data sizes as well as time consumed to train/predict.
DVC has been used to track various data processing and model training stages. This ensures both repeatability and posterity, as the nature of processing is recorded as are the results. The actual data is stored in an S3-compliant object store which isn't publicly available. Artifacts can be made available upon request to [email protected].