hi-paris/ssml-text2breaks-fr-lora is a LoRA adapter fine-tuned on Qwen2.5-7B to predict natural pause locations in French text by adding symbolic <break/> markers.

This is the first stage of a two-step SSML cascade pipeline for improving French text-to-speech prosody control.

📄 Paper: "Improving Synthetic Speech Quality via SSML Prosody Control"
Authors: Nassima Ould-Ouali, Awais Sani, Ruben Bueno, Jonah Dauvet, Tim Luka Horstmann, Eric Moulines
Conference: ICNLSP 2025
🔗 Demo & Audio Samples: https://hi-paris.github.io/DemoTTS/

🧩 Pipeline Overview

Stage	Model	Purpose
1️⃣	hi-paris/ssml-text2breaks-fr-lora	Predicts natural pause locations
2️⃣	hi-paris/ssml-breaks2ssml-fr-lora	Converts breaks to full SSML with prosody

✨ Example

Input:

Bonjour comment allez-vous aujourd'hui ?

Output:

Bonjour comment allez-vous aujourd'hui ?<break/>

🚀 Quick Start

Installation

pip install torch transformers peft accelerate

Basic Usage

python

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

# Load base model and tokenizer
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B",
    torch_dtype=torch.float16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "hi-paris/ssml-text2breaks-fr-lora")

# Prepare input
text = "Bonjour comment allez-vous aujourd'hui ?"
formatted_input = f"### Task:\nConvert text to SSML with pauses:\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)
result = response.split("### SSML:\n")[-1].strip()
print(result)  # "Bonjour comment allez-vous aujourd'hui ?<break/>"

Production Usage (Recommended)

For production use with memory optimization and full cascade, see our inference repository:

python

from text2breaks_inference import Text2BreaksInference

# Memory-efficient shared model approach
model = Text2BreaksInference()
result = model.predict("Bonjour comment allez-vous aujourd'hui ?")

🔧 Full Cascade Example

python

from breaks2ssml_inference import CascadedInference

# Initialize full pipeline (memory efficient)
cascade = CascadedInference()

# Convert plain text directly to full SSML
text = "Bonjour comment allez-vous aujourd'hui ?"
ssml_output = cascade.predict(text)
print(ssml_output)  
# Output: '<prosody pitch="+2.5%" rate="-1.2%" volume="-5.0%">Bonjour comment allez-vous aujourd'hui ?</prosody><break time="300ms"/>'

🧠 Model Details

Base Model: Qwen/Qwen2.5-7B
Fine-tuning Method: LoRA (Low-Rank Adaptation)
LoRA Rank: 8, Alpha: 16
Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training: 5 epochs, batch size 1 with gradient accumulation
Language: French
Model Size: 7B parameters (LoRA adapter: ~81MB)
License: Apache 2.0

📊 Performance

The model achieves high accuracy in predicting natural pause locations in French text, contributing to improved prosody in text-to-speech synthesis when combined with the second-stage model.

🔗 Resources

Full Pipeline Code: https://github.com/hi-paris/cascading_model
Interactive Demo: Colab Notebook
Stage 2 Model: hi-paris/ssml-breaks2ssml-fr-lora

📖 Paper

This model is part of the work described in:

Improving French Synthetic Speech Quality via SSML Prosody Control

If you use this model, please cite the paper.

@inproceedings{ouali-etal-2025-improving,
    title = "Improving {F}rench Synthetic Speech Quality via {SSML} Prosody Control",
    author = "Ouali, Nassima Ould  and
      Sani, Awais Hussain  and
      Bueno, Ruben  and
      Dauvet, Jonah  and
      Horstmann, Tim Luka  and
      Moulines, Eric",
    editor = "Abbas, Mourad  and
      Yousef, Tariq  and
      Galke, Lukas",
    booktitle = "Proceedings of the 8th International Conference on Natural Language and Speech Processing (ICNLSP-2025)",
    month = aug,
    year = "2025",
    address = "Southern Denmark University, Odense, Denmark",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.icnlsp-1.30/",
    pages = "302--314"
}

📜 License

Apache 2.0 License (same as the base Qwen2.5-7B model)

nassimaODL/ssml-text2breaks-fr-lora

🗣️ French Text-to-Breaks LoRA Model

🧩 Pipeline Overview

✨ Example

🚀 Quick Start

Installation

Basic Usage

Production Usage (Recommended)

🔧 Full Cascade Example

🧠 Model Details

📊 Performance

🔗 Resources

📖 Paper

📜 License

nassimaODL/
ssml-text2breaks-fr-lora