Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Predicting depression from acoustic features of speech using a Convolutional Neural Network.
| Date | Stars |
|---|---|
| 2026-07-31 | 319 |
| 2026-08-03 | 319 |
| 2026-08-04 | 319 |
| 2026-08-06 | 319 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# *DepressionDetect*
This effort addresses an automated device for detecting depression from acoustic features in speech. The tool is aimed at lowering the barrier of entry in seeking help for potential mental illness and supporting medical professionals' diagnoses.
Early detection and treatment of depression is essential in promoting remission, preventing relapse, and reducing the emotional burden of the disease. Current diagnoses are primarily subjective, inconsistent across professionals, and expensive for the individual who may be in dire need of help. Additionally, early signs of depression are difficult to detect and quantify. These early signs have a promising potential to be quantified by machine learning algorithms that could be implemented in a wearable artificial intelligence (AI) or home device.
Automatic Depression Detection (ADD) is a relatively nascent topic that first appeared in 2009. *DepressionDetect* presents a novel approach focusing on two aspects that receive scant research attention: class imbalance and data representation (feature extraction).
## Table of Contents
1. [Dataset](#dataset)
2. [Acoustic Features of Speech](#acoustic-features-of-speech)
* [Segmentation](#segmentation-code)
* [Feature Extraction](#feature-extraction-code)
3. [Convolutional Neural Networks](#convolutional-neural-networks)
* [Class Imbalance](#class-imbalance-code)
* [Model Architecture](#model-architecture-code)
* [Training the Model](#training-the-model)
* [Results](#results)
4. [Donate Your Data](#donate-your-data-code)
5. [Future Directions](#future-directions)
For a code walkthrough, see the [src](https://github.com/kykiefer/depression-detect/tree/master/src) folder.
## Dataset
All audio recordings and associated depression metrics were provided by the [DAIC-WOZ Database](http://dcapswoz.ict.usc.edu/), which was compiled by USC's Institute of Creative Technologies and released as part of the 2016 Audio/Visual Emotional Challenge and Workshop ([AVEC 2016](http://sspnet.eu/avec2016/)). The dataset consists of 189 sessions, averaging 16 minutes, between a participant and virtual interviewer called Ellie, controlled by a human interviewer in another room via a "[Wizard of Oz](https://en.wikipedia.org/wiki/Wizard_of_Oz_experiment)" approach. Prior to the interview, each participant completed a psychiatric questionnaire ([PHQ-8](http://patienteducation.stanford.edu/research/phq.pdf)), from which a binary "truth" classification (depressed, not depressed) was derived.
A representative transcribed interview excerpt is seen below:
> **Ellie:** Who’s someone that’s been a positive influence in your life?
> **Participant:** Uh my father.
> **Ellie:** Can you tell me about that?
> **Participant:** Yeah, he is a uh. He’s a very he’s a man of few words. And uh he's very calm. Slow to anger. And um very warm very loving man. Responsible. And uh he’s a gentleman has a great sense of style and he’s a great cook.
<img alt="Virtual interview with Ellie" src="images/interview_with_ellie.png" width='400'>
<sub><b>Figure 1: </b> Virtual interview with Ellie. </sub>
## Acoustic Features of Speech
While some emotion detection research focuses on the semantic content of audio signals in predicting depression, I decided to focus on the [prosodic](http://clas.mq.edu.au/speech/phonetics/phonology/intonation/prosody.html) features, which have also been found to be promising predictors of depression. Prosodic features can be generally characterized by a listener as pitch, tone, rhythm, stress, voice quality, articulation, intonation, etc. Encouraging features in research include sentence length and rhythm, intonation, fundamental frequency, and Mel-frequency cepstral coefficients ([MFCCs](https://en.wikipedia.org/wiki/Mel-frequency_cepstrum)).<sup>[2](#references)</sup>
### Segmentation ([code](https://github.com/kykiefer/depression-detect/blob/master/src/data/segmentation.py))
The first step in analyzing Excerpt of 22,410 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:803e530435e6af70, llm:Repository description and README: 'Predicting depression from acoustic features of speech using a Convolutional Neural Network.' Topics: acoustic-features, cnn, convolutional-neural-networks, depression, spectrogram, speech. Focus on feature extraction, CNN model, training, class imbalance, dataset DAIC-WOZ.
matched fp:803e530435e6af70, llm:Repository description and README: 'Predicting depression from acoustic features of speech using a Convolutional Neural Network.' Topics: acoustic-features, cnn, convolutional-neural-networks, depression, spectrogram, speech. Focus on feature extraction, CNN model, training, class imbalance, dataset DAIC-WOZ.
matched fp:803e530435e6af70, llm:Repository description and README: 'Predicting depression from acoustic features of speech using a Convolutional Neural Network.' Topics: acoustic-features, cnn, convolutional-neural-networks, depression, spectrogram, speech. Focus on feature extraction, CNN model, training, class imbalance, dataset DAIC-WOZ.