Embedded Speech Recognition System (AVR + DTW + MFCC-Lite)

System Architecture

High-Level Pipeline

flowchart LR
A[Microphone Input] --> B[ADC Sampling - audio.c]
B --> C[SRAM Buffer - sram.c]
C --> D[Frame Segmentation]
D --> E[Feature Extraction - dsp.c]
E --> F[UART Streaming]
F --> G[Host PC Tools]
G --> H[DTW Classification]
H --> I[Predicted Word]

Loading

---

Firmware modules

Module	Description
audio.c	ADC sampling + timer control
dsp.c	Feature extraction (MFCC-lite, energy, ZCR, centroid)
fix_fft.c	Fixed-point FFT for spectral analysis
dtw.c	Dynamic Time Warping matching engine
ext_int.c	External interrupt handling
uart.c	UART communication driver
lcd.c	LCD display driver
sram.c	External SRAM control logic

Host side tool chain

Script	Purpose
features_stream.py	Receives live feature vectors from AVR
features_train.py	Trains DTW templates and generates dtw_templates.h
inference.py	Validates AVR predictions vs host model

External SRAM Architecture

The system uses external SRAM for audio buffering due to limited AVR RAM.

Hardware Interface Shared 8-bit data/address bus Two 74373 latch registers (address demultiplexing) SRAM read/write control via bus switching

sequenceDiagram
participant MCU
participant Latch1
participant Latch2
participant SRAM

MCU->>Latch1: Output low address byte
MCU->>Latch2: Output high address byte
MCU->>SRAM: Set address complete

alt Write
MCU->>SRAM: Write data via bus
else Read
MCU->>SRAM: Enable output buffer
SRAM->>MCU: Return data
end

Loading

Audio Acquisition Modes

Stream Mode

by typing in "Stream" inside the uart, the adc samples are streamed directly via uart

Used for debugging the mic and real-time monitoring using the uart_rec.py

DSP mode

the default mode, can also be switched back using dsp

Samples stored in external SRAM (~8 KB buffer), Frames extracted after recording through dumping into the audio buffer

Feature extraction

Each frame is converted into an 8-dimensional feature vector:

• Short-Time Energy
• Zero-Crossing Rate (ZCR)
• Spectral Centroid
• MFCC-lite coefficients

trims silent frames through checking its energy

Classification DTW

Each word is stored as a feature template

Incoming sequence is compared against all templates inside dtw_template.h

Lowest DTW distance determines prediction