AAC Encoder / Decoder – Multimedia Systems Project

This project implements an AAC-like audio encoder and decoder in Python, developed for the Multimedia Systems course.

The system progressively builds three encoding levels:

• Level 1: SSC + MDCT Filterbank
• Level 2: + TNS (Temporal Noise Shaping)
• Level 3: + Psychoacoustic Model + Quantizer + Huffman Coding

The objective is to achieve maximum compression with controlled perceptual distortion, following the core principles of the AAC standard.

---

Full Pipeline (Level 3)

Encoder

WAV (time domain)
    ↓
SSC (Frame Type Selection)
    ↓
FilterBank (MDCT + Windowing)
    ↓
TNS (Frequency-domain prediction)
    ↓
Psychoacoustic Model (SMR computation)
    ↓
Quantizer (Non-uniform quantization)
    ↓
Huffman Encoding
    ↓
.mat compressed structure

Decoder

.mat
    ↓
Huffman Decoding
    ↓
Inverse Quantizer
    ↓
Inverse TNS
    ↓
Inverse FilterBank (IMDCT)
    ↓
Overlap-Add
    ↓
WAV (time domain)

---

Implemented Modules

1) SSC – Sequence Segmentation Control

Frames are classified into:

• OLS – Only Long Sequence
• ESH – Eight Short Sequence
• LSS – Long Start Sequence
• LPS – Long Stop Sequence

Attack detection is performed using a first-order high-pass IIR filter and energy evaluation over 8 segments of 128 samples each.

This mechanism prevents pre-echo artifacts during transient regions.

---

2) Filter Bank – MDCT

• Frame size: 2048 samples
• Overlap: 50% (hop size = 1024)
• Window type: KBD (Kaiser-Bessel Derived)

For ESH frames:

• Central 1152 samples are selected
• Split into 8 overlapping subframes of 256 samples

The MDCT is fully invertible and satisfies the Princen–Bradley condition for perfect reconstruction.

---

3) TNS – Temporal Noise Shaping

• LPC order: p = 4
• Band-wise normalization
• Quantized prediction coefficients
• Stability check
• FIR filtering in the MDCT domain

Purpose:

• Reduce pre-echo
• Redistribute quantization noise temporally
• Improve perceptual quality

---

4) Psychoacoustic Model

The psychoacoustic stage computes:

• FFT per frame/subframe
• Predictability (using two previous frames)
• Band energies
• Spreading function
• Tonality index
• Masking threshold
• SMR (Signal-to-Mask Ratio)

The SMR determines how much quantization noise is allowed per critical band.

---

5) Non-Uniform Quantizer

AAC-style non-uniform quantization:

S(k) = sign(X(k)) ⌊ |X(k) · 2^(−a/4)|^(3/4) ⌋

Dequantization:

X̂(k) = sign(S(k)) · |S(k)|^(4/3) · 2^(a/4)

Iterative Scalefactor Adjustment

Scalefactors are increased while:

• Pe(b) < T(b)
• |a(b+1) − a(b)| ≤ 60

The process stops when:

• All bands satisfy masking constraints, or
• Maximum iterations are reached

---

6) Huffman Coding

• Automatic codebook selection for MDCT symbols
• Codebook 11 forced for scalefactors
• ESC mode support

Per frame metrics recorded:

• bits_S
• bits_sfc
• nonzero coefficients
• maxAbs
• selected codebook
• quantizer iterations

---

Experimental Results

Level 1 & Level 2

Level	SNR (dB)
1	254.03629
2	254.03630

Practically lossless (only floating-point rounding errors).

---

Level 3 (example – max_iter = 60)

• SNR: 10.8 dB
• Bitrate: 177.02 kbps
• Compression Ratio: 8.68×

Effect of max iterations

Max Iter	SNR (dB)	Bitrate (kbps)	Compression
40	24.71	372.59	4.12×
45	22.20	354.03	4.34×
50	18.82	255.18	6.02×
55	14.73	243.57	6.31×
60	10.80	177.02	8.68×
65	7.68	206.07	7.45×
70	6.44	207.86	7.39×
75	5.97	209.72	7.32×
100	5.97	214.14	7.17×
250	5.97	218.00	7.05×

An interesting behavior appears beyond 60 iterations, where bitrate starts increasing again due to large scalefactor differences increasing Huffman cost.

---

How to Run

From the project root:

python run_level_1.py
python run_level_2.py
python run_level_3.py

Or with full path:

python path/to/run_level_3.py

---

Outputs

• output_levelX.wav – Decoded audio
• coded_level3_metrics.csv – Per-frame encoding metrics

---

Academic Context

Developed as part of the Multimedia Systems course assignment.
Implements a simplified yet structurally faithful AAC-like codec.

---