Update README.md

README.md

@@ -7,4 +7,95 @@ sdk: static
 pinned: false
 ---
 
-Edit this `README.md` markdown file to author your organization card.
+<p align="center">
+  <img src="https://raw.githubusercontent.com/mathisescriva/gilbert_landing_page/main/assets/images/img/logo_gilbert.png" width="180" />
+</p>
+
+# Gilbert AI  
+**Sovereign Speech Intelligence Research Lab**
+
+---
+
+## 1. Mission
+
+Gilbert AI develops sovereign, high-accuracy automatic speech recognition (ASR) systems designed for institutional, educational, and enterprise environments.
+
+The core objectives include:
+
+- achieving state-of-the-art ASR performance for **French**, including long-form, spontaneous, multi-speaker, and accented speech;
+- ensuring robustness across heterogeneous conditions (telephone bandwidth, online conferencing, real-world noise);
+- maintaining strong multilingual capabilities while specializing in French;
+- delivering reproducible and scientifically-rigorous evaluation frameworks;
+- enabling deployment on sovereign infrastructures with strict data governance constraints.
+
+---
+
+## 2. Research Themes
+
+### **2.1 High-precision French ASR**
+Tailored to:
+
+- meetings and long-form speech  
+- lectures and institutional discourse  
+- spontaneous multi-speaker environments  
+- regional and international accents  
+
+### **2.2 Multilingual preservation**
+Although optimized for French, models are evaluated to maintain strong cross-lingual performance.
+
+### **2.3 Robustness, frugality, and domain adaptation**
+Research includes:
+
+- noise augmentation and channel variability  
+- 8 kHz telephony adaptation  
+- low-compute inference  
+- distillation for edge deployment  
+
+### **2.4 Benchmarking and dataset engineering**
+Gilbert designs domain-specific datasets and benchmark suites for evaluating ASR in conditions representative of operational environments.
+
+---
+
+## 3. Current Model Family
+
+### **Gilbert-FR-Source (2025)**
+Baseline model fine-tuned on curated high-quality French corpora; foundation for all domain-specific variants.
+
+### Upcoming research releases:
+
+- **Gilbert-FR-Longform-v1** — extended-speech and meeting optimization  
+- **Gilbert-FR-Accents-v1** — regional and international accent specialization  
+- **Gilbert-FR-Telephone-v1** — 8 kHz telephony channel adaptation  
+- **Gilbert-Edu-ASR-v1** — education-specific speech model  
+- **Gilbert-Multilingual-v1** — French-centric multilingual enhancement  
+
+---
+
+## 4. Baseline Performance (WER)
+
+| Dataset | WER |
+|--------|------|
+| MLS (FR) | **3.98%** |
+| Common Voice v13 (FR) | **7.28%** |
+| VoxPopuli (FR) | **8.91%** |
+| Fleurs (FR) | **4.84%** |
+| African-accent French | **4.20%** |
+
+These results position Gilbert-FR among the strongest open-source ASR models available for the French language.
+
+---
+
+## 5. Research Principles
+
+1. **Reproducibility**  
+2. **Scientific rigor**  
+3. **Transparent evaluation**  
+4. **Sovereign infrastructure**  
+5. **Operational relevance**  
+
+---
+
+## 6. Contact
+
+- Website: **https://gilbert-assistant.fr**  
+- Contact: **mathis@lexiapro.fr**