Need help with emotion-recognition-using-speech?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

145 Stars 92 Forks MIT License 59 Commits 4 Opened issues


Building and training Speech Emotion Recognizer that predicts human emotions using Python, Sci-kit learn and Keras

Services available


Need anything else?

Contributors list

# 17,565
51 commits

Speech Emotion Recognition


  • 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. ## Requirements
  • Python 3.6+ ### Python Packages
  • 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

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

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
for training and
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 in `create
script in
` 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 "20190616125714_happy.wav" will be parsed automatically as happy)

Emotions available

There are 9 emotions available: "neutral", "calm", "happy" "sad", "angry", "fear", "disgust", "ps" (pleasant surprise) and "boredom".

Feature Extraction

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

Grid search results are already provided in

folder, but if you want to tune various grid search parameters in
, you can run the script
This may take several hours to complete execution, once it is finished, best estimators are stored and pickled in

Example 1: Using 3 Emotions

The way to build and train a model for classifying 3 emotions is as shown below: ```python from 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


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())

Test score: 0.8148148148148148 Train score: 1.0 ```

Determining the best model

In order to determine the best model, you can by:

# loads the best estimators from `grid` folder that was searched by GridSearchCV in ``,
# and set the model to the best in terms of test score, and then train it
# get the determined sklearn model name
print(rec.model.__class__.__name__, "is the best")
# get the test accuracy score for the best estimator
print("Test score:", rec.test_score())


MLPClassifier is the best
Test Score: 0.8958333333333334


Just pass an audio path to the

method as shown below: ```python

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/tessravdess/validation/Actor25/25010101mob_sad.wav"))

Prediction: neutral Prediction: sad ```

Example 2: Using RNNs for 5 Emotions

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
# get the accuracy
# predict angry audio sample
prediction = deeprec.predict('data/validation/Actor_10/03-02-05-02-02-02-10_angry.wav')
print(f"Prediction: {prediction}")


Prediction: angry
Predicting probabilities is also possible (for classification ofc):
{'angry': 0.8502438, 'sad': 1.15252915e-05, 'neutral': 8.986728e-05, 'ps': 0.14671412, 'happy': 0.0029406736}

Confusion Matrix

print(deeprec.confusion_matrix(percentage=True, labeled=True))


              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

Algorithms Used

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) ### Regressors
  • SVR
  • RandomForestRegressor
  • GradientBoostingRegressor
  • KNeighborsRegressor
  • MLPRegressor
  • BaggingRegressor
  • Recurrent Neural Networks (Keras)


You can test your own voice by executing the following command:

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

for more information.
python --help
Output: ``` usage: [-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"

## Plotting Histograms
This will only work if grid search is performed.
from emotion_recognition import plot_histograms
# plot histograms on different classifiers


A Histogram shows different algorithms metric results on different data sizes as well as time consumed to train/predict.

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.