🎉 macOS Support Arrives!

mojo-audio now runs natively on macOS with blazing performance on Apple Silicon!

✨ What's New

Platform Support:

• ✅ macOS Apple Silicon (M1/M2/M3/M4)
• ✅ macOS Intel (x86_64)
• ✅ Linux x86_64

Performance (Apple Silicon M-series):

• 30s audio: 8.34ms (3,598x realtime)
• 10s audio: 2.81ms (3,562x realtime)
• 1s audio: 0.40ms (2,530x realtime)

vs Python librosa:

• 1.75x faster on 30s audio
• 4.5x faster on 10s audio
• 1.6x faster on 1s audio

🔧 Technical Improvements

• Fixed Mojo 0.26.2+ API compatibility
• Switched to OpenBLAS for cross-platform BLAS support
• Inlined FFI types for reliable shared library builds
• All tests passing on macOS and Linux

📦 Installation

macOS:

curl -L https://github.com/itsdevcoffee/mojo-audio/releases/download/v0.2.0/mojo-audio-v0.2.0-macos-arm64.tar.gz | tar xz
sudo cp libmojo_audio.dylib /usr/local/lib/
sudo cp mojo_audio.h /usr/local/include/

Build from source:

pixi install
pixi run build-ffi-optimized

See README.md for detailed build instructions.

mojo-audio v0.1.0 - Release Notes

Release Date: January 22, 2026

High-performance mel spectrogram preprocessing library in Mojo that beats Python's librosa by 1.5-3.6x on short/medium audio.

🎉 What's New

First Public Release

Complete Whisper-compatible audio preprocessing pipeline built from scratch in Mojo, achieving production-ready performance and beating Python's librosa at all audio durations.

Performance Highlights

Key advantages:

• 1.5-3.6x faster on short/medium audio (most common use case)
• Far more consistent (5-10% variance vs librosa's 22-39%)
• ~1100x realtime throughput on 30s audio
• Zero dependencies, pure Mojo implementation

📦 Release Assets

Pre-built Binaries (FFI)

mojo-audio-v0.1.0-linux-x86_64.tar.gz (15KB)

Pre-built shared library for Linux x86_64:

• libmojo_audio.so - Optimized shared library (26KB, -O3)
• mojo_audio.h - C header file
• INSTALL.md - Installation and usage guide

Use from:

• ✅ C/C++
• ✅ Rust
• ✅ Python (ctypes/cffi)
• ✅ Go (cgo)
• ✅ Any language with C FFI support

mojo-audio-ffi-examples.tar.gz (3.1KB)

FFI usage examples:

• demo_c.c - Complete C example
• demo_rust.rs - Complete Rust example
• Makefile - Build system for examples

Source Code

• Source code (zip) - Auto-generated by GitHub
• Source code (tar.gz) - Auto-generated by GitHub

✨ Features

Core Audio Processing

• Mel Spectrogram Pipeline - Complete Whisper-compatible preprocessing
• Window Functions - Hann and Hamming with SIMD optimization
• FFT Operations - Radix-2/4 iterative FFT + true RFFT for real signals
• STFT - Parallelized short-time Fourier transform across CPU cores
• Mel Filterbank - Sparse-optimized triangular filters (80 or 128 bands)
• Normalization - Multiple modes (Whisper, min-max, z-score, raw)

FFI Support

• Zero-overhead C API - Same performance as native Mojo
• Type-safe interface - Proper error handling and memory management
• Comprehensive examples - C, Rust, Python usage samples
• Easy installation - Pre-built binaries for Linux x86_64

Testing & Benchmarking

• 17 test cases - All passing
• Robust methodology - Deterministic signals, warmup, outlier exclusion
• Comparison scripts - Back-to-back mojo-audio vs librosa
• Web UI - Interactive benchmark visualization

🚀 Quick Start

Native Mojo

git clone https://github.com/itsdevcoffee/mojo-audio.git
cd mojo-audio
pixi install
pixi run demo-mel

FFI (C/Rust/Python)

# Download and install pre-built library
wget https://github.com/itsdevcoffee/mojo-audio/releases/download/v0.1.0/mojo-audio-v0.1.0-linux-x86_64.tar.gz
tar xzf mojo-audio-v0.1.0-linux-x86_64.tar.gz
cd linux-x86_64
sudo cp libmojo_audio.so /usr/local/lib/
sudo cp mojo_audio.h /usr/local/include/
sudo ldconfig
# Test it
wget https://github.com/itsdevcoffee/mojo-audio/releases/download/v0.1.0/mojo-audio-ffi-examples.tar.gz
tar xzf mojo-audio-ffi-examples.tar.gz
gcc demo_c.c -lmojo_audio -lm -o demo
./demo

🔧 Requirements

Native Mojo

• Mojo 0.26.1 or later
• pixi package manager
• Linux (primary platform)

FFI (Pre-built Binaries)

• Linux x86_64
• glibc 2.31 or later
• No Mojo installation required!

🎯 Whisper Compatibility

Fully compatible with OpenAI Whisper models:

• ✅ Sample rate: 16kHz
• ✅ FFT size: 400
• ✅ Hop length: 160 (10ms frames)
• ✅ Mel bands: 80 (v2) or 128 (v3)
• ✅ Output shape: (n_mels, ~3000) for 30s
• ✅ Normalization: Whisper-compatible mode

📚 Documentation

• README - Overview and usage
• FFI Integration Guide - Complete FFI documentation
• CHANGELOG - Detailed change log
• CONTRIBUTING - Contribution guidelines

🏆 Achievements

• 🥇 Beats librosa at all durations (1.1-3.6x faster)
• 🚀 17-68x speedup from naive to optimized implementation
• ⚡ ~1100x realtime throughput on 30s audio
• 📊 Far more consistent than librosa (5-10% vs 22-39% variance)
• ✅ 100% from scratch in Mojo - no black-box dependencies
• 🎓 Complete learning resource with educational examples

🔬 Technical Highlights

9 Major Optimizations

1. Iterative FFT (3.0x) - Cache-friendly Cooley-Tukey
2. Twiddle Precompute (1.7x) - Pre-computed rotation factors
3. Sparse Mel Filterbank (1.24x) - Skip zero-weight bins
4. Twiddle Caching (2.0x) - Reuse across frames
5. Float32 Precision (1.07x) - 2x SIMD width
6. True RFFT (1.43x) - Exploits real signal symmetry
7. Parallelization (1.3-1.7x) - Multi-core frame processing
8. Radix-4 FFT (1.1-1.2x) - Optimized butterflies
9. Compiler Optimization (1.2-1.5x) - -O3 flag

Total: 17-68x faster than naive implementation!

🐛 Known Issues

None reported. This is the initial stable release.

🤝 Contributing

Contributions are welcome! See CONTRIBUTING.md for guidelines.

Areas where we'd love help:

• Additional FFI language bindings (Go, Julia, Zig, Swift)
• ARM platform support and optimization
• Additional audio features (MFCC, CQT, etc.)
• Documentation improvements

📝 License

MIT License - See LICENSE

🔗 Links

• Repository: https://github.com/itsdevcoffee/mojo-audio
• Issues: https://github.com/itsdevcoffee/mojo-audio/issues
• Discussions: https://github.com/itsdevcoffee/mojo-audio/discussions

📊 Benchmark Methodology

All benchmarks use improved methodology:

• Signal: Deterministic chirp (20Hz-8000Hz, reproducible)
• Warmup: 5 runs (JIT stabilization)
• Iterations: 20 (statistical significance)
• Analysis: Outlier exclusion, mean ± std dev
• Comparison: Back-to-back mojo-audio vs librosa

Run benchmarks yourself:

pixi run bench-compare # Full comparison
pixi run bench-stable 5 # Stable (5 runs, median)

Built with Mojo 🔥 | Faster than Python | Production-Ready