This research studies whether a language model fine-tuned on a person's writing can reproduce their written style. “Voice” here means writing style, not speech or audio. The paper evaluates QLoRA adapters using authorship attribution, controls that suppress topic words, memorization checks, and comparisons against prompting-based baselines.

The authors report that low-rank adapters capture much of the measured stylometric signal on their test data. They also examine memorization, the effects of training duration, cross-domain behavior, and whether generated text is more stereotypically attributable than an author's own writing. See the paper for the experimental design, findings, and limitations; the results should not be interpreted as a guarantee that the same behavior generalizes to other models or writers.

paper/paper.pdf — named-author preprint.
paper/ — LaTeX source, bibliography, figures, tables, and the ACL style files used for the manuscript.
code/ — scripts for data preparation, training, generation, stylometric evaluation, analysis, and figures.
requirements.txt — packages used in the experiments; see the installation caveat below.
configs/ and results/ — locations for files generated by the scripts.

This repository does not include the raw source corpora, full generated-text outputs, trained adapters, or the original full run outputs. The paper includes a short qualitative example with source and generated text. It is a research code and paper release, not a packaged dataset or model checkpoint. Those omissions limit independent verification and exact reproduction from this repository alone.

Setup

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
python -m pip install 'transformers==4.53.3'
mkdir -p data results/generations results/adapters configs

Dependency caveat: The original environment used mlx-lm==0.31.3 with transformers==4.53.3. The project's requirements note that this combination overrides the declared mlx-lm dependency floor; a resolver may report a conflict. It has not been validated on all systems. Some training scripts invoke the interpreter at .venv/bin/python, so create the environment at that path.

Data preparation

Obtain the corpora separately and comply with their respective terms. The scripts do not download them automatically.

Download and extract the UCI Reuter 50/50 corpus, placing the C50train and C50test directories directly under data/.

From the repository root, prepare the original eight-author newswire subset:

PYTHONHASHSEED=0 python code/prepare_data.py
python code/validate_stylometry.py

For the expanded newswire study, code/prepare_more_authors.py prepares additional authors. For cross-domain runs, code/prepare_domain.py blog expects blog XML files under data/raw/blogs/, and code/prepare_domain.py enron expects an Enron maildir under data/raw/maildir/; the CMU Enron corpus is distributed separately. Review applicable permissions and privacy considerations before obtaining, processing, or redistributing those texts.

The preparation scripts use Python's built-in hash() for part of the shuffling. Setting PYTHONHASHSEED makes that behavior repeatable for a given environment; the seed used for the original run was not recorded, so this does not establish that the original splits will be recovered.

Running the baseline pipeline

From the repository root, after preparing data:

python code/train_grid.py
python code/generate.py baselines
python code/generate.py lora
python code/evaluate.py

These commands are the initial newswire pipeline, not a complete reproduction of all results in the paper. phase1_train.py, phase2_train.py, phase3_baselines.py, phase3_scale.py, analysis_phase1.py, analysis_domains.py, memorization_v2.py, openworld_verify.py, and contamination_probe.py contain further runs and analyses. Read their docstrings and inspect their inputs before execution. Some expect outputs from earlier stages and will take substantial time and disk space. llm_judge.py calls an external API and requires credentials and paid API access; do not commit credentials or private text.

The scripts generate configuration files and outputs locally. Existing result files may cause scripts to skip completed conditions. Original splits, versioned dependencies, checkpoints, and complete output artifacts are not provided, so exact numerical replication has not been verified.

Building the paper

With Tectonic installed:

cd paper
tectonic -X compile main.tex

The repository includes the manuscript's generated numbers, bibliography, and figures; rebuilding the PDF does not rerun the experiments.

Citation

@misc{shams2026voice,
  title  = {How Much of You Fits in a Low-Rank Matrix? Quantifying the Stylometric Fidelity of Parameter-Efficient Voice Cloning},
  author = {Shams, Shakeeb and Mitash, Yukt},
  year   = {2026}
}

Responsible use and limitations

The work is intended to help measure writing-style fidelity and memorization. Personalized generation can expose or imitate someone's writing without their consent; use appropriate consent and review generated outputs for memorized text. The experiments use particular corpora, model sizes, and attribution judges; the paper's limitations and ethics sections describe the scope of the conclusions.

License

The original project code is released under the MIT License. Third-party datasets, models, and manuscript style files retain their own terms; their inclusion or use is not covered by the code license. The underlying corpora and model weights are not distributed here.

Downloads last month: -

Total file size:

591 kB

How Much of You Fits in a Low-Rank Matrix?

Contents