🎧 Audio Processing Workflow

🧰 Tools Used

1. 📥 Reading the Audio File

• We use librosa to load and process audio files.
• The sample is a 4-second audio clip sampled at 22050 Hz.
• Time-domain waveform plot:

mmmm.mov

2. 🔍 Voice Activity Detection (VAD) & Clipping

• VAD detects segments with human speech.

• Steps:

  • Normalize signal amplitude to [-1, 1]
  • Apply a threshold to filter speech vs noise
  • Clip the audio to keep only speech segments

VAD Output Plot:

• Red: VAD mask
• Blue: Original signal
• Selected range: Retained for clipping

Cropped Signal Plot:

2.vad_croped.mov

3. 📈 Pre-Emphasis

• Enhances high frequencies before analysis.

• Helps:

  1. Balance spectrum
  2. Avoid numerical issues in FFT
  3. Improve SNR

Filter Equation:

$$ y(t) = x(t) - \alpha x(t-1),\quad \alpha = 0.97 $$

Plot after Pre-emphasis:

3.preempha.mov

4. 🧩 Splitting the Audio

• Audio is split into 10ms frames using a sliding window.

Frame Sequence Visualization:

Single Frame Example:

5. 🪟 Hann Window Function

• Reduces spectral leakage by smoothing frame edges.

Window Function:

$$ w(n) = 0.5 \left(1 - \cos\left({2\pi nN}\right)\right),\quad 0 \leq n \leq N $$

Windowed Frame Sequence:

Windowed Single Frame:

6. 🔄 Time → Frequency Domain (FFT)

• We apply Fast Fourier Transform (FFT) to convert from time to frequency domain.

Time Domain:

Frequency Domain:

7. 🌈 Spectrogram

• A spectrogram displays how frequency content changes over time.

Axes:

• X-axis: Time
• Y-axis: Frequency
• Color: Amplitude (Power in dB)

Spectrogram:

8. 🔍 MFCC Feature Extraction

• MFCCs capture perceptual characteristics of audio relevant to human hearing.

• Widely used in speech recognition.

• Frequency perception is:

  • Linear < 1kHz
  • Logarithmic > 1kHz

Axes:

• X-axis: Time
• Y-axis: MFCC coefficients (12 in our case)

MFCC Plot:

9. 🔄 Reconstructing Audio from MFCC

• Audio can be partially reconstructed from MFCCs with acceptable quality loss.

4.reconstructed.mov

📜 License

MIT License © 2025-PRESENT wudi