Fat Llama

fat_llama is a Python package for upscaling audio files to FLAC or WAV formats using advanced audio processing techniques. It utilizes CUDA-accelerated calculations to enhance audio quality by upsampling and adding missing frequencies through FFT (Fast Fourier Transform), resulting in richer and more detailed audio.

Features

• Upscale MP3 files to high-quality FLAC format.
• Band-limited (FFT-domain) interpolation and iterative soft thresholding (IST) for enhanced audio processing, with no bandwidth extension beyond the original recording's frequency ceiling.
• LMS adaptive filtering, auto-scaling amplitude adjustment, and normalization.
• Supports GPU-accelerated processing with CuPy.
• Claude is used for coding and maintainance of this package, but not for generating audio.

Requirements

• CUDA capable GPU

(Note: For cpu verison please look at https://pypi.org/project/fat-llama-fftw/)

Installation

Install via pip:

pip install fat-llama

Note: This version works with CUDA 13 (tested against 13.3).

Further need CUDA & CuPy properly installed: https://docs.cupy.dev/en/stable/install.html

Also, requires ffmpeg: https://support.audacityteam.org/basics/installing-ffmpeg

Note to install on older versions of CUDA and CuPy. You will need to download specific versions and install locally.

• cupy version - https://github.com/bkraad47/fat_llama/tree/v-0.1.3---cupy
• cupy-cuda11x version - https://github.com/bkraad47/fat_llama/tree/v-0.1.3---cupy-cuda11x
• cupy-cuda12x version - https://github.com/bkraad47/fat_llama/tree/v-1.1.0

To install locally:

git clone <target_url>
cd fat_llama
pip install .

Usage

Example Usage

You can run the example provided in example.py:

from fat_llama.audio_fattener.feed import upscale

# Example call to the method
upscale(
    input_file_path='input_test.mp3',
    output_file_path='output_test.flac',
    source_format='mp3',
    target_format='flac',
    max_iterations=300,
    threshold_value=0.6,
    target_bitrate_kbps=1400,
    toggle_normalize=True,
    toggle_autoscale=True,
    toggle_adaptive_filter=True
)

Function Parameters

• input_file_path (str): Path to the input audio file. Mandatory.
• output_file_path (str): Path to the output processed audio file. Mandatory.
• source_format (str): Format of the input audio file (e.g., 'mp3', 'wav', 'ogg', 'flac').
• target_format (str): Format of the output audio file (e.g., 'flac', 'wav'). Default is 'flac'.
• max_iterations (int): Maximum number of iterations for IST. Default is 300.
• threshold_value (float): Threshold value for IST. Default is 0.6.
• target_bitrate_kbps (int): Used to derive the upscale factor relative to the source file's own bitrate; must fall within the valid range for target_format (800-1411 kbps for flac, 800-6444 kbps for wav). Default is 1411. Note this is not a promise about the output file's real bitrate — the output is always written as uncompressed PCM at an upsampled sample rate, so its actual bitrate will be substantially higher.
• toggle_normalize (bool): Whether to normalize the audio. Default True.
• toggle_autoscale (bool): Whether to autoscale the audio based on the original audio. Default True.
• toggle_adaptive_filter (bool): Whether to apply LMS adaptive filtering. Default True.

Running the Example

To run the example, execute the following command:

python example.py

This will upscale the MP3 file specified in the example and produce a FLAC file with full processing.

Spectrogram Results

Audio Quality Scores

Generated by the test-fat-llama skill's audio-quality-checker subagent — updated each run, not hand-edited.

How it works

Algorithm Explanation

The upscaling process involves several steps:

1. Reading Audio File: The audio file is read, and the audio samples are extracted along with the sample rate and bitrate.
2. Calculating Upscale Factor: The upscale factor is derived from target_bitrate_kbps relative to the source file's own bitrate, then clamped so the resulting sample rate never exceeds a realistic consumer playback ceiling (192 kHz) and never goes below the original sample rate — the derivation no longer produces arbitrarily large factors (e.g. 250-300+ kHz output sample rates) that the pipeline's own Nyquist cutoff (step 8) would have made mostly silent anyway.
3. Upscaling Channels: The audio channels are upscaled using a bandlimited FFT-domain interpolation (zero-padding the spectrum then inverse-transforming) rather than naive sample repetition, so the extra samples don't introduce spectral imaging above the original signal's Nyquist frequency.
4. Iterative Soft Thresholding (IST): IST is applied to enhance the audio by adding missing frequencies. This process uses FFT to transform the signal into the frequency domain, apply a threshold to keep significant frequencies, and then inverse transform back to the time domain, repeated for max_iterations rounds. A per-iteration synthetic harmonic-reconstruction term was tried across cycles 2-3 (content-derived frequency, peak-scaled amplitude), but was found (cycle 4) to be computed from a single whole-buffer FFT, so for any real multi-second upscaled channel its "content-derived" frequency was actually static across the whole track — a constant, audible, non-source tone that collapsed measured dynamic range in quiet passages. It has been removed rather than revised a fourth time; IST now performs the plain FFT/threshold/IFFT round trip described above with no synthetic tone added on top.
5. Scaling Amplitude: The amplitude of the upscaled audio is scaled to match the original.
6. Normalizing Audio: The audio is normalized to the range -1 to 1.
7. Adaptive Filtering: A block-adaptive LMS filter with a short decorrelation delay refines the normalized signal, adapting its coefficients based on the signal's own short-term predictability. Tap weights update once per block of samples (rather than once per individual sample) using the block-averaged gradient, which is what makes enabling this stage practical at typical upscaled sample counts.
8. Original-Nyquist Cutoff: An unconditional final FFT-domain lowpass removes any spectral content above the original source file's Nyquist frequency, guaranteeing the upscale never synthesizes or leaves behind content beyond the original recording's real bandwidth — upscaling improves precision and headroom within that bandwidth, it does not extend it.
9. Writing FLAC File: The processed audio is written to a FLAC file.

Why FFT and IST?

FFT (Fast Fourier Transform) is used to transform the audio signal into the frequency domain. This allows for the identification and manipulation of specific frequency components. By applying a threshold in the frequency domain, we can keep significant frequencies and discard noise and add it to our upscaling data to add detail to upscaling frequencies.

The report titled "Fast Sparse Fourier Transformations for NMR Spectroscopy" by Badruddin Kamal, supervised by Thomas Huber and Alastair Rendall, 2015, provides a comprehensive understanding of sparse representations and their applications in signal processing. IST leverages the concepts from this report to add missing frequencies and enhance the audio quality by making it more detailed and rich. This is particularly useful in upscaling audio where some frequencies might be missing or congested.

Test Audio Source

ericzo - beyond link(https://soundcloud.com/ericzomusic/free-electro-trap-anthem-beyond)

Changelog

The full, current changelog is maintained in CHANGELOG.md — see it for the latest entries going forward. The history below is kept for reference.