GitHub

https://github.com/Earwigmoth10/stuttering-detection-classifier.git

title: SpeakFlow AI emoji: colorFrom: blue colorTo: teal sdk: docker app_port: 5000 pinned: false

SpeakFlow AI

AI-Powered Speech Stuttering Detection System

SpeakFlow AI analyzes uploaded or recorded speech audio and detects whether it contains normal speech or stuttering patterns, using MFCC feature extraction and a Random Forest classifier.

Features

Audio upload & live in-browser recording
ML-based stutter detection (40 MFCC features + Random Forest)
Confidence & fluency scoring on every prediction
Downloadable PDF report of each analysis
Searchable analysis history
User accounts (sign up, log in, edit profile, change password)
Admin dashboard — user management, analytics charts, CSV export
Dark mode, responsive UI

Frontend

Tech Stack

Frontend: HTML5, CSS3, Vanilla JavaScript, Chart.js, jsPDF, Font Awesome Backend: Python, Flask, Flask-CORS, Librosa, Scikit-learn, NumPy, Joblib

How It Works

User uploads or records a WAV/MP3 file
Audio is preprocessed (noise reduction, normalization, silence trimming)
40 MFCCs are extracted using Librosa
Feature vector is classified by a trained Random Forest model
App returns Normal / Stutter prediction with a confidence score

Dataset

This model was trained using publicly available speech datasets from the following sources:

Stuttering Speech Dataset

UCLASS Stuttered Speech Clips (SEP-28k Format)
Source: https://www.kaggle.com/datasets/vudominhgiang/uclass-stuttered-speech-clips-sep-28k-format

Fluent (Normal) Speech Dataset

Mozilla Common Voice
Source: https://commonvoice.mozilla.org/

The datasets were preprocessed and combined to create a binary classification dataset for distinguishing between Fluent Speech and Stuttering Speech.

Note: The datasets are not redistributed in this repository. Please download them from their respective official sources and ensure compliance with their licenses before use.

Running This Space

This Space runs a Flask backend that serves the ML prediction API. On startup it will be available at the URL shown in the Space's embedded app window.

Backend endpoint: /api/predict (POST, multipart audio upload)
Health check: /

Note: This app was originally built to run with a separate local frontend (index.html + http.server) talking to 127.0.0.1:5000. When deployed here, the frontend's API base URL points at this Space's backend instead of localhost.

Admin Access

An admin account is configured in the login flow to access the admin dashboard (user management, analytics, CSV export).

Demo project notice: User accounts, sessions, and analysis history are stored in browser localStorage, not a real database, and passwords are not hashed. This project is for academic/demo purposes only — not intended for production use with real user data.

Model Details

Feature extraction: 40 MFCCs averaged over time (Librosa)
Classifier: Random Forest (Scikit-learn)
Classes: Normal speech vs. Stuttered speech
Training data: 8,000+ labeled audio samples

Author

Laiba Aamir (reawigmoth)

Downloads last month: -; Downloads are not tracked for this model. How to track