Interspeech submission

Supplementary materials for our submitted paper to Interspeech 2025

---

1. Repo structure:

This repository provides supplementary materials for our submission entitled "Towards High-Quality LLM-Based Data for French Spontaneous Speech Simplification: an Exo-Refinement Approach" for the next edition of Interspeech. The main goal is to provide all the codes and prompt templates to generate reproducible results.

To this end, we include:

• The designed code along with the different generative prompting templates to implement the iterative generation-evaluation workflow and thus produce synthetic simplifications.
• The set of collected expert-based simplifications, that allowed to evaluate the quality of the generated outputs under different automatic metrics.

interspeech-2025/
├── README.md
├── data/
│   ├── experts/ 
│   │   └── expert_simplifications_all.tsv  # Collected reference simplifications
│   └── orfeo/
│       └── conversations/    # Set of 240 sentences from orfeo-cefc set
├── generate_data_mistral.py  # Script to run Mistral-based LLMs
├── generate_data_openai.py   # Script to run OpenAI-based LLMs
├── modules/
│   ├── functions_mistral.py
│   ├── functions_openai.py
│   └── templates/
│       ├── evaluation.py     # Evaluation prompts
│       ├── generation.py     # Generation prompts
│       └── topeval.py        # Top-evaluator prompt
└── requirements.txt

---

2. Installation:

Prerequisites:

• Python 3.10.8 or above.
• Obtain a MistralAI API key: https://console.mistral.ai/api-keys/ (alternatively, run LLMs locally).
• Obtain an OpenAI API key: https://openai.com/index/openai-api/.

$ git clone git@github.com:lormaechea/interspeech-2025.git
$ cd interspeech-2025
$ pip install -e requirements.txt

---

3. How to use:

$ python generate_data_mistral.py # To generate Mistral-based synthetic simplifications
$ python generate_data_openai.py  # To generate OpenAI-based synthetic simplifications

4. Citation

@inproceedings{ormaechea-2025-speech-simplification,
    title = {Towards High-Quality LLM-Based Data for French Spontaneous Speech Simplification: an Exo-Refinement Approach},
    author = {Lucía Ormaechea, Nikos Tsourakis, Pierrette Bouillon, Benjamin Lecouteux and Didier Schwab},
    booktitle = {Proc. Interspeech 2025},
    year = {2025},
    location = {Rotterdam, Netherlands},
    url = {https://www.isca-archive.org/interspeech_2025/ormaechea25_interspeech.html},
}