Depression Detection Based on Speech activity in this project we are trying to observe and study depression effect on speech mostly on speech activity, meaning number of silences in speech and length of this silences first we need to exctract silences from speech, we are going to do this wtith our SAD model then after extracting silences, we perform two-tailed t-test then we are going to use Spectrograms exctracted from speech to detect depression
This database is part of a larger corpus, the Distress Analysis Interview Corpus (DAIC) (Gratch et al.,2014), that contains clinical interviews designed to support the diagnosis of psychological distress conditions such as anxiety, depression, and post-traumatic stress disorder. These interviews were collected as part of a larger effort to create a computer agent that interviews people and identifies verbal and nonverbal indicators of mental illness (DeVault et al., 2014). Data collected include audio and video recordings and extensive questionnaire responses; this part of the corpus includes the Wizard-of-Oz interviews, conducted by an animated virtual interviewer called Ellie, controlled by a human interviewer in another room. Data has been transcribed and annotated for a variety of verbal and non-verbal features.
This share includes 189 sessions of interactions ranging between 7-33min (with an average of 16min). Each session includes transcript of the interaction, participant audio files, and facial features. For more details please refer to the documentation
you can download this dataset at Download
first we use VUV(voiced/unvoiced) provided by the dataset itself for our target label, which is not a tottaly accurate assumption
we exctracted MFFCs and their deltas from interviews, and use them as input for our model
model is a hybrid model using both CNN and GRUs BatchNormalization after CNNs for training speedup using Dropout after each layer of GRUs and the before the output layer we used a GlobalAveragePooling Layer
we used tf.Keras for defining model and google colab for Training, and we achived the result in blew table
| _ | Train_set | Validation_set | Test_set |
|---|---|---|---|
| AUC | 0.9457 | 0.9432 | 0.9290 |
first we cleaned interviews audio files, to contain only Participants Voice then used Audiomentations a publicly available python library, to augment Train_set then splited each audio file to 2-Secs samples and exctracted Spectrogram for each of them we used a Conv_1d Model for this purpose and managed to achive 0.86 recall on the test set, which means we could mostly identify depressed people