Fixed-Point MP3 Encoder

A high-efficiency, fixed-point MP3 encoder optimized for embedded systems and resource-constrained environments.

Table of Contents

• Quick Start
• Overview
• Audio Quality
• Performance
• Historical Context
• Applications
• Building
• Usage
• Project Structure
• Documentation
• License

Quick Start

# Build encoder and tools
make

# Encode an audio file
./pem_encode input.wav output.mp3

# Run quick test
make quick-test

# Run full test suite
make test

Overview

This encoder implements MPEG-1 Layer III (MP3) audio compression using entirely fixed-point arithmetic. Originally developed for the first generation of portable media players, it achieves real-time encoding on processors as modest as the ARM7TDMI at 74 MHz with only 40 KB of RAM.

Key Features

• Fixed-point arithmetic: No floating-point unit required
• Minimal memory footprint: ~40 KB RAM, ~70 KB ROM
• Real-time capable: Designed for 1x realtime on 74 MHz ARM7
• CBR/VBR/ABR modes: Constant, variable, and average bitrate encoding
• Quality presets: LAME-style -V 0 through -V 9 presets
• Bitrate range: 32-320 kbps
• Sample rates: 32 kHz, 44.1 kHz, 48 kHz (MPEG-1 Layer III)
• ID3 tags: ID3v1 and ID3v2.3 metadata support
• Format: MPEG-1 Layer III stereo (mono input auto-converted)
• Modern platform support: macOS ARM64/x86_64, Linux

Copyright

Copyright (C) 1998-2025 Mark Phillips. All rights reserved.

Original algorithm: Segher Boessenkool (1998-2002)
Embedded integration: Interactive Objects, Inc. (2001)
Modern platform adaptation: Mark Phillips (2025)

Audio Quality

Objective Measurements

Comparison of original LAME-encoded MP3 versus re-encoding through the PEM fixed-point encoder at 192 kbps:

Metric	Original (LAME)	PEM Encoder	Difference
Mean Volume	-23.000 dB	-23.200 dB	0.200 dB
Peak Volume	-4.400 dB	-4.800 dB	0.400 dB
Duration	29.989 s	29.944 s	0.045 s
File Size	703 KB	701 KB	0.284%

Quality Characteristics

The encoder produces perceptually acceptable output with the following trade-offs:

Aspect	Characteristic
Psychoacoustic model	Simplified energy-based (vs. full FFT)
Stereo mode	Mid-side only (no intensity stereo)
Block switching	Not implemented (long blocks only)
Pre-emphasis	Not implemented
Best performance	128-256 kbps

Performance

Computational Overhead

Platform	Speed	Notes
Apple M1 (3.2 GHz)	680x realtime	30 seconds encoded in 0.044 s
EP7312 (74 MHz ARM7)	~1x realtime	Original target platform
Estimated MIPS	~60 MIPS	For realtime at 128 kbps

Memory Requirements

Component	Size	Description
Code segment	98 KB	Executable instructions
Data segment	82 KB	Initialized data
SRAM working set	40 KB	Runtime state and buffers
ROM tables	45 KB	Pre-computed constants

The 40 KB working set fits entirely within the EP7312's 80 KB on-chip SRAM, eliminating external memory access during encoding.

Encoding Symmetry

Traditional MP3 encoding is 3-10x more expensive than decoding. This encoder achieves near-symmetric performance through:

1. Simplified psychoacoustic model: Energy-based vs. FFT-based
2. Pre-computed tables: All trigonometric values in ROM
3. Single-pass quantization: Binary search vs. iterative optimization
4. Fixed block size: No adaptive block switching

Historical Context

Target Platform: Cirrus Logic EP7312

The encoder was designed for the EP7312 ARM720T-based system-on-chip, popular in first-generation portable media players (1998-2002).

Specification	EP7312
Processor	ARM720T @ 74 MHz
Architecture	ARMv4T (Thumb)
Instruction cache	8 KB
Data cache	None
On-chip SRAM	80 KB
External memory	16-bit SDRAM

The EP7312 datasheet is included at docs/ep7312.pdf.

Commercial Deployment

This encoder technology was deployed in:

• Portable MP3 players (1999-2003)
• Pocket PC voice recording applications
• Automotive entertainment systems
• Embedded audio logging devices

Applications

Beyond Audio Compression

The fixed-point FFT and spectral analysis techniques used in this encoder have applications in scientific instrumentation, particularly where low-power, real-time frequency analysis is required.

Insect Species Identification

A notable application is optical insect detection, as demonstrated by Mullen et al. (2016). The system uses the same spectral analysis principles:

Species	Wing Beat Frequency
Mosquitoes (Culicidae)	300-600 Hz
House flies (Musca domestica)	100-200 Hz
Honey bees (Apis mellifera)	130-250 Hz
Citrus psyllid (D. citri)	187 +/- 26 Hz

How it works:

1. A tracking laser illuminates a flying insect
2. A photodiode measures oscillating light intensity as wings beat
3. Signal conditioning filters noise (high-pass DC removal, 2 kHz low-pass)
4. FFT via Welch's method transforms the time-domain signal to frequency
5. Spectral signature matches against species database

The Welch overlapped periodogram method enables species classification with:

• Solar-powered field deployment
• Microcontroller-class processors
• Same fixed-point FFT code as this encoder

The research paper is included at docs/oe-24-11-11828.pdf.

Other Fixed-Point FFT Applications

Domain	Application
Biomedical	ECG rhythm analysis, EEG classification
Industrial	Vibration monitoring, motor diagnostics
Environmental	Seismic detection, acoustic monitoring
Agricultural	Crop disease spectral analysis

Building

Requirements

• GCC or Clang compiler
• Make
• ffmpeg (optional, for testing)

Build Commands

# Build encoder and all tools
make

# Build encoder only
make pem_encode

# Build with debug symbols
make CFLAGS_EXTRA="-g"

# Build with sanitizers
make CFLAGS_EXTRA="-fsanitize=address,undefined"

# Show all targets
make help

Make Targets

Target	Description
all	Build encoder and tools (default)
pem_encode	Build encoder only
tools	Build analysis tools only
quick-test	Quick encode test (2 second sine wave)
test	Run full test suite
roundtrip-test	Quality comparison with mp3-fixed samples
generate-samples	Generate test WAV files only
encode-samples	Generate and encode all test samples
clean	Remove build artifacts (keeps samples)
distclean	Remove all generated files
install	Install to /usr/local/bin

Usage

Basic Encoding

# Encode at default 128 kbps CBR
./pem_encode input.wav output.mp3

# Encode at specific bitrate
./pem_encode -b 192 input.wav output.mp3

# Show statistics after encoding
./pem_encode -s input.wav output.mp3

# Quiet mode (no progress bar)
./pem_encode -q input.wav output.mp3

VBR and ABR Modes

# VBR with quality preset (0=best, 9=smallest)
./pem_encode -V 2 input.wav output.mp3

# ABR targeting 128 kbps average
./pem_encode --abr 128 input.wav output.mp3

# VBR with bitrate constraints
./pem_encode -V 4 --vbr-min 96 --vbr-max 256 input.wav output.mp3

Psychoacoustic Tuning

The encoder provides fine-grained control over the psychoacoustic model for optimizing audio quality.

# Use a tuning preset
./pem_encode --preset quality input.wav output.mp3
./pem_encode --preset voice input.wav output.mp3

# Adjust individual parameters
./pem_encode --ath 50 --temporal 30 --gain -5 input.wav output.mp3

Tuning Presets

Preset	Description	Best For
default	Balanced quality and size	General purpose
quality	Less aggressive masking	High fidelity
speed	More aggressive masking	Fast encoding
voice	Optimized for speech	Podcasts, audiobooks
music	Balanced for music	Songs, albums
bass	Enhanced low frequencies	Electronic, hip-hop
transparent	Near-transparent quality	Archival, masters

Tuning Parameters

Parameter	Range	Default	Description
--ath N	0-100	0	ATH (Absolute Threshold of Hearing) sensitivity
--temporal N	0-100	50	Frame-to-frame scalefactor smoothing
--gain N	-20 to +20	0	Global gain offset (negative = higher quality)

Stereo Mode Options

Control how stereo audio is encoded for optimal quality or file size.

# Force specific stereo mode
./pem_encode --stereo-mode adaptive input.wav output.mp3

# Adjust MS stereo threshold
./pem_encode --ms-threshold 30 input.wav output.mp3

# Control stereo width
./pem_encode --stereo-width 80 input.wav output.mp3

Stereo Modes

Mode	Description
auto	Automatic selection (default)
stereo	Simple L/R stereo encoding
joint	Joint stereo (MS encoding)
ms	Mid-side only, no switching
adaptive	Adaptive MS with per-coefficient decisions

Stereo Parameters

Parameter	Range	Default	Description
--ms-threshold N	0-100	50	MS threshold (lower = wider stereo image)
--stereo-width N	0-100	50	Stereo width (0=narrow, 100=wide)

Quality Presets

Preset	Bitrate Range	Quality
-V 0	220-320 kbps	Transparent
-V 2	170-210 kbps	High (default VBR)
-V 4	140-185 kbps	Good
-V 6	115-150 kbps	Acceptable
-V 9	65-85 kbps	Low bitrate

ID3 Tags

# Add metadata tags
./pem_encode --title "Song Name" --artist "Artist" --album "Album" input.wav output.mp3

# Control tag versions
./pem_encode --id3v2 --no-id3v1 input.wav output.mp3  # ID3v2 only
./pem_encode --id3v1 --no-id3v2 input.wav output.mp3  # ID3v1 only

Supported Bitrates

32, 40, 48, 56, 64, 80, 96, 112, 128, 160, 192, 224, 256, 320 kbps

Input Requirements

Parameter	Requirement
Format	WAV (PCM)
Channels	1 (mono) or 2 (stereo)
Sample rate	32000, 44100, or 48000 Hz
Bit depth	16-bit

Analysis Tools

# Generate test signals
./tools/wav_generator --sine 440 5 test.wav
./tools/wav_generator --sweep 20 20000 10 sweep.wav
./tools/wav_generator --noise 5 noise.wav
./tools/wav_generator --multitone 5 multi.wav

# Analyze WAV file
./tools/wav_analyzer input.wav

# Inspect MP3 frame structure
./tools/mp3_info output.mp3

Project Structure

mp3-fixed-encoder/
├── pemlib/              # Core encoder library
│   ├── include/         # Public headers
│   │   └── fpmp3.h      # Main API header
│   └── src/             # Implementation
│       ├── codec.c      # Main encoder orchestration
│       ├── polyphase.c  # Polyphase analysis filter bank
│       ├── hybrid.c     # MDCT and aliasing butterfly
│       ├── psy.c        # Psychoacoustic model
│       ├── psy_tuning.c # Psychoacoustic tuning parameters
│       ├── psy_tuning.h # Tuning interface header
│       ├── quant.c      # Quantization and rate control
│       ├── huffman.c    # Huffman encoding tables
│       ├── out.c        # Bitstream formatting
│       ├── reservoir.c  # Bit reservoir management
│       ├── vbr.c        # VBR/ABR bitrate control
│       ├── id3tag.c     # ID3v1/v2 metadata tags
│       └── ro.c         # ROM constant tables
├── src/                 # CLI application
│   └── main.c           # Command-line interface
├── tools/               # Analysis utilities
│   ├── wav_generator.c  # Test signal generator
│   ├── wav_analyzer.c   # WAV file analyzer
│   └── mp3_info.c       # MP3 frame inspector
├── docs/                # Technical documentation
│   ├── iso11172-3/      # MP3 specification excerpts
│   ├── 02_mp3_algorithm.md
│   ├── 03_fixed_point_math.md
│   ├── 04_optimization_generic.md
│   ├── 05_encoder_optimizations.md
│   ├── 06_fft_applications.md
│   ├── ep7312.pdf       # Target processor datasheet
│   └── oe-24-11-11828.pdf  # Insect detection paper
├── Makefile
├── README.md
├── TODO.md
└── LICENSE

Documentation

Technical documentation is available in the docs/ directory:

Document	Description
05_encoder_optimizations.md	Encoder-specific optimizations and design decisions
06_fft_applications.md	FFT applications beyond audio (insect detection)
iso11172-3/	ISO MPEG-1 Audio specification excerpts
ep7312.pdf	Cirrus Logic EP7312 processor datasheet
oe-24-11-11828.pdf	Mullen et al. optical insect detection paper

License

This software is provided under a non-commercial license. See LICENSE for full terms.

Summary:

• Personal and educational use permitted
• Commercial use requires separate license
• No redistribution without permission
• Attribution required

For commercial licensing inquiries, contact Mark Phillips through the repository.

References

1. ISO/IEC 11172-3:1993 - Information technology - Coding of moving pictures and associated audio for digital storage media at up to about 1.5 Mbit/s - Part 3: Audio

2. Mullen, E.R., et al. (2016). "Optical sensors for the detection of flying insects." Optics Express, 24(11), 11828.

3. Cirrus Logic EP7312 Data Sheet - 32-bit ARM720T Core Processor

4. Davis, P. (1998). "Fixed-point DSP for Audio Applications." Embedded Systems Programming.