README.md

metadata

license: apache-2.0
base_model: Qwen/Qwen2.5-7B
library_name: peft
tags:
  - text-to-speech
  - ssml
  - qwen2.5
  - lora
  - peft
language:
  - en
  - fr
pipeline_tag: text-generation

🗣️ ssml-text2breaks-fr-lora

ssml-text2breaks-fr-lora is a LoRA adapter built on top of Qwen/Qwen2.5-7B, trained to predict symbolic pause markers (e.g., #250, #500) in raw French text. These symbolic tags indicate appropriate prosodic boundaries for speech synthesis systems.

This model is the first stage in the cascaded pipeline presented in:

"Improving French Synthetic Speech Quality via SSML Prosody Control"
Nassima Ould-Ouali, Éric Moulines – ICNLSP 2025 (Springer LNCS, accepted)

It is designed to be followed by ssml-break2ssml-fr-lora, which converts symbolic markers into valid SSML tags.

---

🧩 Pipeline Overview

Stage	Model Name	Description
1️⃣	ssml-text2breaks-fr-lora	Predicts symbolic pause markers such as #250, #500
2️⃣	ssml-break2ssml-fr-lora	Converts symbolic markers into <break time="..."/> SSML tags

---

✨ Example

Input:

Bonjour je m'appelle Bertrand Perier. Je suis avocat à la cour.

Output

Bonjour#250 je m'appelle Bertrand Perier.#500 Je suis avocat à la cour.

🧠 Model Details

• Base Model: Qwen/Qwen2.5-7B
• Fine-tuning Method: LoRA (Low-Rank Adaptation)
• LoRA Rank: 8
• LoRA Alpha: 16
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Training Epochs: 5
• Batch Size: 1 (with gradient accumulation)
• Learning Rate: 3e-4

🚀 How to run the code

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

# Load base model and tokenizer
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "jonahdvt/qwen-ssml-lora")

# Prepare input
instruction = "Convert text to SSML with pauses:"
text = "Hello, how are you today? I hope everything is going well."
formatted_input = f"### Task:\n{instruction}\n\n### Text:\n{text}\n\n### SSML:\n"

# Generate
inputs = tokenizer(formatted_input, return_tensors="pt").to(model.device)
with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
        temperature=0.7,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
ssml_output = response.split("### SSML:\n")[-1]
print(ssml_output)

Citation

If you use this model in your research, please cite:

@inproceedings{ould-ouali2025_improving,
  title     = {Improving Synthetic Speech Quality via SSML Prosody Control},
  author    = {Ould-Ouali, Nassima and Sani, Awais and Bueno, Ruben and Dauvet, Jonah and Horstmann, Tim Luka and Moulines, Eric},
  booktitle = {Proceedings of the 8th International Conference on Natural Language and Speech Processing (ICNLSP)}, % TODO: vérifier l'intitulé exact utilisé par la conf
  year      = {2025},
  pages     = {XX--YY},   % TODO
  publisher = {—},        % TODO 
  address   = {—}         % TODO
}

License

This model is released under the Apache 2.0 license, same as the base Qwen2.5-7B model.