AVIL is a C++20 command-line tool for macOS that detects whether a WAV file is secretly a transcoded MP3, and visualizes live audio frequency spectra directly in your terminal — backed by an AVX2-accelerated FFT engine.
When downloading tracks from early 2000s DJ CDs, I often found that files advertised as high quality sounded surprisingly poor. This tool was born out of that frustration — it detects whether an audio file is genuinely high quality or just a low-bitrate MP3 (ripped from sources like YouTube) that has been re-encoded into a lossless format to fake quality.
MP3 transcoding detection — loads a WAV file, runs an FFT-based spectral analysis across multiple frames, and flags files whose high-frequency energy drops off sharply above 16 kHz (a characteristic artifact of MP3 encoding). Reports a verdict based on how many frames exceed the cutoff threshold.
Real-time spectrum visualizer — renders a live, colour-coded frequency spectrogram and stereo volume meters in the terminal using FTXUI. Works both for audio files and microphone input.
Microphone passthrough with feedback suppression — captures microphone input, suppresses acoustic feedback via frequency-domain thresholding, and plays processed audio back to the output device in real time.
1. Skip the first few seconds (silence / intro fades)
2. For each analysis frame:
a. Downmix multichannel audio to mono
b. Compute RMS — skip frames that are too silent
c. Apply a Hann window to reduce spectral leakage
d. Run the Cooley-Tukey FFT (AVX2-accelerated)
e. Compare high-frequency energy (>16 kHz) against
mid-frequency energy (10–16 kHz)
f. Flag the frame if the ratio falls below threshold
3. Report "probably MP3" if >70% of valid frames were flagged
The key insight is that MP3 encoders introduce a hard cutoff in the upper frequency range. Genuine lossless files retain energy well past 16 kHz; upscaled files do not.
The FFT is a Cooley-Tukey radix-2 DIT implementation with an AVX2-optimised butterfly stage. When four or more butterflies remain in a block, the engine processes them simultaneously using 256-bit SIMD registers — each holding four complex numbers (8 × 32-bit floats). The complex multiplication (a+bi)(c+di) is carried out with _mm256_moveldup_ps / _mm256_movehdup_ps for real/imaginary duplication and _mm256_addsub_ps for the final combine, giving a theoretical 4× throughput improvement over the scalar path.
The inverse FFT (ifft) is implemented by conjugating the input, running the forward FFT, then conjugating and scaling the output — no separate code path is needed.
# Microphone capture + real-time visualizer
./build/avil
# File analysis (detection + playback + visualizer)
./build/avil path/to/track.wavPress q or Ctrl+C to exit the TUI.
Upscaling analysis: Probably the file has been upscaled
#CUTOFFS/TOTAL = 18/22
Upscaling analysis: The file doesn't look upscaled
#CUTOFFS/TOTAL = 1/24
- macOS 10.15+
g++with C++20 and AVX2 supportmake
git clone https://github.com/giorgiogamba/avil.git
cd avil
# Install dependencies (builds a local lib/ folder)
make install_dependencies
# Compile
make build/avil
# Run
./build/avilTo clean build artefacts:
make cleanThe AVX2 butterfly path processes 4 complex numbers per instruction cycle, delivering roughly 4× throughput over the scalar fallback:
| Operation | Generic | AVX2 Optimised | Speedup |
|---|---|---|---|
| FFT (1024 samples) | 245 µs | 62 µs | 3.95× |
| Full file analysis | 1.2 s | 340 ms | 3.53× |
| Real-time processing | 87% CPU | 24% CPU | 3.63× |
Measured on MacBook Pro (2020), Intel Core i7-1068NG7.
| Library | Purpose |
|---|---|
| PortAudio | Audio I/O (microphone capture and playback) |
| libsndfile | Audio file decoding (WAV, AIFF, …) |
| FTXUI | Terminal UI (spectrogram, volume meters) |
avil/
├── src/
│ ├── main.cpp # Entry point, TUI, stream callbacks, detection logic
│ ├── fft.hpp # Cooley-Tukey FFT with AVX2 butterfly optimisation
│ ├── types.h # Shared data structures (StreamCallbackData, FileCallbackData)
│ └── constants.h # Tuning parameters (buffer sizes, frequency bounds, thresholds)
├── makefile
├── run.sh
└── README.md
- Improve upscaling detection heuristics (entropy, spectral flatness)
- Batch-mode analysis (scan a directory of files)
- FLAC, AAC, and OGG input support
- Cross-platform audio I/O (Linux / Windows)
- GPU-accelerated FFT path
Contributions are welcome. Good places to start:
- Tighten the detection algorithm (false-positive/negative analysis)
- Add support for additional audio formats via libsndfile
- Improve cross-platform portability of the audio I/O layer
Please open an issue before starting significant work so we can coordinate.
MIT — see LICENSE for details.