How benekashetty013/kannada-ipa-asr works
This repository is a small Kannada speech-recognition project: it records labeled audio, extracts MFCC features, trains a CNN, and presents predictions in a Streamlit interface. The graph distinguishes those stages and the stored dataset, feature arrays, model weights, and prediction log. The README documents the overall workflow; the excerpts show feature extraction and training behavior, plus parts of inference and result rendering. App internals and the exact runtime wiring between artifacts are only partially excerpted, so those links are marked unverified where appropriate.
Main entry points: app.py Streamlit Interface · record_dataset.py Audio Recorder · dataset Labeled WAVs · extract_features.py MFCC Extraction
Components 13
Audio Data
- Audio Recorder
dataset/record_dataset.py - Labeled WAVs
dataset
Model Pipeline
- MFCC Extraction
extract_features.py - Feature Arrays
features - CNN Training
train_model.py - Model Weights
model/best_model.pt
Recognition UI
- Streamlit Interface
app.py - Audio Preprocessing
app.py - CNN Inference
app.py - Word Details
app.py - Acoustic Analysis
app.py - Prediction Log
results/predictions_log.csv
Other components
- Speaker
Connections 14
- Speaker → Audio Recorder: speaks word
dataset/record_dataset.py - Audio Recorder → Labeled WAVs: saves recordings
dataset/record_dataset.py - Labeled WAVs → MFCC Extraction: extracts features
extract_features.py - MFCC Extraction → Feature Arrays: writes arrays
extract_features.py - Feature Arrays → CNN Training: trains from
train_model.py - CNN Training → Model Weights: saves weights
train_model.py - Speaker → Streamlit Interface: provides audio
README.md - Streamlit Interface → Audio Preprocessing: prepares audio
app.py - Audio Preprocessing → CNN Inference: provides tensor
app.py - Model Weights → CNN Inference: loads model
- CNN Inference → Streamlit Interface: returns predictions
app.py - Streamlit Interface → Word Details: looks up word
app.py - Streamlit Interface → Acoustic Analysis: renders spectrogram
app.py - Streamlit Interface → Prediction Log: logs prediction
app.py