Streaming Digit classifier (lightweight)
Real-time Spoken Digit Classification (0-9)
A real-time audio digit classification system that recognizes spoken numbers (0-9) through live microphone streaming. Features multiple classification approaches including TTS APIs, Fourier analysis, MFCC, and MEL features with performance benchmarking and inference time tracking.
How to Use
▶
1. Click "Start Recording" and clearly say a digit (0-9)
2. Select a processing method from the cabinets below
3. Watch the real-time audio visualization
4. Compare inference times and accuracy across methods
5. Use the Robustness toggle to test with noise
Audio Input Monitor
?
Ready to record...
Duration: 0.0s
Duration: 0.0s
Audio Robustness Settings
0.0
Select Processing Pipeline
MFCC + Dense NN
Mel Frequency Cepstral Coefficients
Predicted Digit:
?
Confidence:
--
Inference Time:
0.0ms
Testing Acc:
98.52%
Mel CNN (2D)
Mel Spectrogram Convolutional Neural Network
Predicted Digit:
?
Confidence:
--
Inference Time:
0.0ms
Testing Acc:
97.22%
Raw CNN (1D)
Raw Waveform Convolutional Neural Network
Predicted Digit:
?
Confidence:
--
Inference Time:
0.0ms
Testing Acc:
91.30%
External API
Whisper API (Disabled)
Predicted Digit:
N/A
Confidence:
--
Inference Time:
N/A
Status:
Disabled
Model Architecture & Training Metrics
MFCC + Dense NN
| Architecture | 13 MFCC → Dense(128) → Dense(64) → Dense(10) |
|---|---|
| Test Accuracy | 98.52% |
| Validation Accuracy | 97.89% |
| Parameters | 10,314 |
| Training Time | 3.2 minutes |
| Inference Time | ~1-2ms |
Mel CNN (2D)
| Architecture | 2D CNN → MaxPool → Dense(128) → Dense(10) |
|---|---|
| Test Accuracy | 97.22% |
| Validation Accuracy | 96.45% |
| Parameters | 45,782 |
| Training Time | 8.7 minutes |
| Inference Time | ~3-5ms |
Raw CNN (1D)
| Architecture | 1D CNN → Conv1D → GlobalMaxPool → Dense(10) |
|---|---|
| Test Accuracy | 91.30% |
| Validation Accuracy | 89.67% |
| Parameters | 28,954 |
| Training Time | 12.1 minutes |
| Inference Time | ~5-8ms |
External API
| Model | Whisper (Hugging Face) |
|---|---|
| Parameters | ~39M (External) |
| Language | English Speech Recognition |
| Connection | HTTPS API Call |
| Latency | ~1-3 seconds |
| Test Accuracy | Variable (Network dependent) |
Raw Spectrogram (Dropped)
| Architecture | STFT → 2D CNN → Dense Layers |
|---|---|
| Status | Not Implemented |
| Reason | High Dimensionality |
| Issue | Memory intensive (~64k features) |
| Alternative | Mel-scale features used instead |
| Estimated Performance | Similar to Mel CNN but slower |
MFCC + SVM (Alternative)
| Architecture | 13 MFCC → Support Vector Machine |
|---|---|
| Status | Not Implemented |
| Estimated Accuracy | 85-90% |
| Advantages | Lightweight, Fast Training |
| Parameters | ~1000 Support Vectors |
| Inference Time | ~0.5ms (Estimated) |
Performance Monitor
Total Predictions
0
Fastest Method
-
Session Time
00:00
API Latency
-- ms
