STM32F407 Real-Time Audio DSP System

Real-time audio processing system using STM32F407 Discovery board with overlap-add FFT-based spectral filtering.

🎯 Features

• Real-time audio processing at 48kHz sample rate
• 512-point FFT with 50% overlap-add for frequency-domain filtering
• Low-latency pipeline: ~10.7ms total delay (512 samples)
• Spectral filtering: Configurable low-pass, high-pass, band-pass, and notch filters
• Hardware acceleration: CMSIS-DSP library optimized for ARM Cortex-M4F
• DC offset removal: Adaptive tracking for clean audio input

📋 Hardware Requirements

• STM32F407VG Discovery Board
• Audio Input: Analog signal to PA1 (ADC1_IN1)
• Audio Output: CS43L22 DAC via I2S3 (on-board)
• External Components: None required (uses on-board codec)

🔧 Technical Specifications

Parameter	Value
MCU	STM32F407VG (168MHz ARM Cortex-M4F)
Sample Rate	48 kHz
ADC Resolution	12-bit
DAC Resolution	16-bit stereo
FFT Size	512 samples
Hop Size	256 samples (50% overlap)
Processing Latency	~10.7 ms
Window Function	Hann (COLA-compliant)

🏗️ System Architecture

Audio Input (PA1)(Initial version is on MAX4466)
    ↓
ADC1 (12-bit, DMA, Timer-triggered @ 48kHz)
    ↓
DC Offset Removal & Float Conversion
    ↓
Overlap-Add FFT Processing (512-point)
    ├─ Hann Windowing
    ├─ FFT (CMSIS-DSP)
    ├─ Spectral Mask Application
    └─ IFFT
    ↓
Float to Q15 Conversion
    ↓
I2S3 DMA (Stereo output)
    ↓
CS43L22 Codec → Headphones/Speakers

🚀 Getting Started

Prerequisites

• STM32CubeIDE (or Keil MDK / IAR EWARM)
• STM32CubeMX (for hardware reconfiguration)
• CMSIS-DSP Library (included in STM32Cube)

Quick Test

1. Connect audio source to PA1 (ADC input)
2. Connect headphones to 3.5mm jack on Discovery board
3. Power on the board
4. You should hear the processed audio output

🎛️ Configuration

Enable Spectral Filtering

Edit audio_processing.c → Init_Spectral_Mask():

Low-pass filter (remove high frequencies above 4kHz):

float cutoff_freq = 4000.0f;
int cutoff_bin = (int)(cutoff_freq * FFT_SIZE / SAMPLE_RATE);
for (int k = cutoff_bin; k <= FFT_SIZE / 2; k++)
{
    spectral_mask[k] = 0.0f;
}

Notch filter (remove 1kHz hum):

float notch_freq = 1000.0f;
int notch_bin = (int)(notch_freq * FFT_SIZE / SAMPLE_RATE);
int notch_width = 2;
for (int k = notch_bin - notch_width; k <= notch_bin + notch_width; k++)
{
    if (k >= 0 && k <= FFT_SIZE / 2)
        spectral_mask[k] = 0.0f;
}

Voice band-pass (300Hz - 3400Hz):

float low_freq = 300.0f;
float high_freq = 3400.0f;
int low_bin = (int)(low_freq * FFT_SIZE / SAMPLE_RATE);
int high_bin = (int)(high_freq * FFT_SIZE / SAMPLE_RATE);
for (int k = 0; k <= FFT_SIZE / 2; k++)
{
    if (k < low_bin || k > high_bin)
        spectral_mask[k] = 0.0f;
}

Adjust Output Gain

Edit audio_processing.h:

#define AUDIO_GAIN  1.0f  // Change to desired gain (e.g., 1.5 for +3dB)

Modify DC Offset Tracking Speed

Edit audio_processing.h:

#define DC_ALPHA  0.01f  // Lower = slower tracking, more stable
                         // Higher = faster tracking, less stable

🧪 Testing

1. Sweep test: Input 20Hz-20kHz sine sweep, verify passband
2. THD measurement: Input 1kHz tone, measure harmonic distortion
3. Latency test: Input impulse, measure output delay with oscilloscope
4. Filter verification: Enable filters, verify frequency response with spectrum analyzer

🐛 Known Issues / Limitations

• Startup transient: Brief ~2ms pop during DC offset convergence
• No anti-aliasing filter: Input signals above 24kHz will alias
• Fixed-point not used: Float32 processing (could optimize for speed)
• No adaptive filtering: Spectral mask is static

🔮 Future Improvements

• Implement adaptive noise reduction (spectral subtraction)
• Add automatic gain control (AGC)
• Port to fixed-point Q15 for lower CPU usage
• Add USB audio class for PC connectivity
• Implement pitch shifting / time stretching

👤 Author

• GitHub: @silanus23
• Email: [email protected]

🙏 Acknowledgments

• STMicroelectronics for HAL drivers and CMSIS-DSP
• ARM for Cortex-M4F DSP extensions and optimized math libraries