Update README.md
Browse files
README.md
CHANGED
|
@@ -24,108 +24,172 @@ tags:
|
|
| 24 |
- research
|
| 25 |
- gilbert
|
| 26 |
---
|
| 27 |
-
# Gilbert-FR-Source
|
| 28 |
|
| 29 |
-
|
| 30 |
|
| 31 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
|
| 33 |
---
|
| 34 |
|
| 35 |
-
## 1.
|
| 36 |
|
| 37 |
-
|
| 38 |
|
| 39 |
-
-
|
| 40 |
-
-
|
| 41 |
-
-
|
| 42 |
-
-
|
| 43 |
-
-
|
| 44 |
|
| 45 |
-
|
|
|
|
| 46 |
|
| 47 |
---
|
| 48 |
|
| 49 |
-
## 2.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
|
| 51 |
-
|
| 52 |
|
| 53 |
-
|
| 54 |
-
|
|
|
|
|
|
|
| 55 |
| MLS (FR) | 3.98 % |
|
| 56 |
| Common Voice FR (v13.0) | 7.28 % |
|
| 57 |
| VoxPopuli (FR) | 8.91 % |
|
| 58 |
| Fleurs (FR) | 4.84 % |
|
| 59 |
| African Accented French | 4.20 % |
|
| 60 |
|
| 61 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 62 |
|
| 63 |
---
|
| 64 |
|
| 65 |
-
##
|
|
|
|
|
|
|
| 66 |
|
| 67 |
-
|
| 68 |
-
|
|
|
|
|
|
|
|
|
|
| 69 |
|
| 70 |
-
|
| 71 |
-
- capacité à modéliser des séquences longues ;
|
| 72 |
-
- pré-entraînement sur de larges corpus multilingues ;
|
| 73 |
-
- forte spécialisation implicite en français observée dans les benchmarks publics ;
|
| 74 |
-
- compatibilité avec les runtimes optimisés (CTranslate2, ONNX Runtime, MLX).
|
| 75 |
|
| 76 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 77 |
|
| 78 |
---
|
| 79 |
|
| 80 |
-
##
|
| 81 |
|
| 82 |
-
|
| 83 |
|
| 84 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 85 |
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
|
|
|
|
|
|
|
|
|
| 90 |
|
| 91 |
---
|
| 92 |
|
| 93 |
-
##
|
|
|
|
|
|
|
| 94 |
|
| 95 |
-
-
|
| 96 |
-
-
|
| 97 |
-
-
|
| 98 |
-
-
|
| 99 |
-
-
|
|
|
|
|
|
|
| 100 |
|
| 101 |
---
|
| 102 |
|
| 103 |
-
##
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 104 |
|
| 105 |
-
|
| 106 |
-
|
| 107 |
|
| 108 |
-
|
| 109 |
-
|
| 110 |
|
| 111 |
-
|
|
|
|
|
|
|
|
|
|
| 112 |
|
| 113 |
---
|
| 114 |
|
| 115 |
-
##
|
|
|
|
|
|
|
| 116 |
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
- Gilbert-FR-Téléphone-v1 (8 kHz, call center, voix compressée)
|
| 120 |
-
- Gilbert-Multilingue-v1 (extension multi-langue)
|
| 121 |
|
| 122 |
-
|
| 123 |
|
| 124 |
---
|
| 125 |
|
| 126 |
-
##
|
| 127 |
|
| 128 |
-
|
| 129 |
-
- Site : https://gilbert-assistant.fr
|
| 130 |
-
- Contact : mathis@lexiapro.fr
|
| 131 |
|
|
|
|
|
|
|
|
|
| 24 |
- research
|
| 25 |
- gilbert
|
| 26 |
---
|
|
|
|
| 27 |
|
| 28 |
+
# Gilbert-FR-Source — Research Baseline for French Automatic Speech Recognition
|
| 29 |
|
| 30 |
+
`Gilbert-FR-Source` is a French automatic speech recognition (ASR) model used as the **research foundation** for the Gilbert project.
|
| 31 |
+
It is designed as an internal scientific baseline enabling controlled experimentation, reproducible evaluation, and rigorous comparison across ASR architectures, datasets, and adaptation methods.
|
| 32 |
+
|
| 33 |
+
This model is not a fine-tuned derivative, but a **curated research anchor** used to support systematic studies in:
|
| 34 |
+
|
| 35 |
+
- domain adaptation,
|
| 36 |
+
- robustness to spontaneous and long-form speech,
|
| 37 |
+
- accented and low-resource linguistic profiles,
|
| 38 |
+
- telephony and bandwidth-constrained speech,
|
| 39 |
+
- multi-speaker and meeting transcription.
|
| 40 |
|
| 41 |
---
|
| 42 |
|
| 43 |
+
## 1. Research Motivation
|
| 44 |
|
| 45 |
+
The Gilbert project aims to build highly specialized ASR systems optimized for:
|
| 46 |
|
| 47 |
+
- professional meeting transcription (hybrid/remote),
|
| 48 |
+
- long-form multi-speaker discourse,
|
| 49 |
+
- institutional environments (education, public sector),
|
| 50 |
+
- constrained audio conditions (telephony, VoIP, low SNR),
|
| 51 |
+
- sociolinguistic diversity (African, Canadian, Belgian and other French accents).
|
| 52 |
|
| 53 |
+
While Whisper Large V3 provides strong baseline performance, its behavior under domain shifts (spontaneous interactions, overlapping speech, degraded microphones) requires systematic study.
|
| 54 |
+
`Gilbert-FR-Source` provides the **frozen starting point** for this line of research, ensuring controlled comparisons between experiments.
|
| 55 |
|
| 56 |
---
|
| 57 |
|
| 58 |
+
## 2. Scientific Goals and Research Questions
|
| 59 |
+
|
| 60 |
+
This model is used to answer a series of research questions:
|
| 61 |
+
|
| 62 |
+
### **Q1. Long-form modeling**
|
| 63 |
+
How does Whisper-L3 behave on meetings lasting 30–120 minutes, with natural topic shifts, interruptions, and pragmatic markers?
|
| 64 |
+
|
| 65 |
+
### **Q2. Accent robustness**
|
| 66 |
+
Which classes of French accents induce the strongest WER degradation?
|
| 67 |
+
How does robustness vary across FLEURS, African French, and Common Voice subsets?
|
| 68 |
+
|
| 69 |
+
### **Q3. Telephony adaptation**
|
| 70 |
+
What is the degradation curve when downsampling to 16 kHz / 8 kHz / μ-law compressed audio?
|
| 71 |
+
|
| 72 |
+
### **Q4. Domain adaptation efficiency**
|
| 73 |
+
What is the marginal gain of targeted fine-tuning on professional meeting datasets (education, administration, healthcare)?
|
| 74 |
+
|
| 75 |
+
### **Q5. Multilingual side-effects**
|
| 76 |
+
To what extent does French fine-tuning affect cross-lingual generalization?
|
| 77 |
+
|
| 78 |
+
These research axes structure the development of future specialized Gilbert models.
|
| 79 |
+
|
| 80 |
+
---
|
| 81 |
|
| 82 |
+
## 3. Benchmark Reference Results
|
| 83 |
|
| 84 |
+
The following WER scores originate from established open benchmarks and serve as a *reference baseline* for future experiments:
|
| 85 |
+
|
| 86 |
+
| Dataset | WER |
|
| 87 |
+
|--------|-----|
|
| 88 |
| MLS (FR) | 3.98 % |
|
| 89 |
| Common Voice FR (v13.0) | 7.28 % |
|
| 90 |
| VoxPopuli (FR) | 8.91 % |
|
| 91 |
| Fleurs (FR) | 4.84 % |
|
| 92 |
| African Accented French | 4.20 % |
|
| 93 |
|
| 94 |
+
These results provide **upper bounds** before targeted fine-tuning.
|
| 95 |
+
Future Gilbert variants will be evaluated using:
|
| 96 |
+
|
| 97 |
+
- internal meeting datasets,
|
| 98 |
+
- domain-specific corpora (administration, higher education, healthcare),
|
| 99 |
+
- accented speech corpora,
|
| 100 |
+
- telephony datasets,
|
| 101 |
+
- long-form evaluation methods (> 1 hour audio).
|
| 102 |
|
| 103 |
---
|
| 104 |
|
| 105 |
+
## 4. Architecture
|
| 106 |
+
|
| 107 |
+
The model is based on the **Whisper Large V3** encoder–decoder architecture, offering:
|
| 108 |
|
| 109 |
+
- large multilingual pretraining,
|
| 110 |
+
- long-context modeling capacity,
|
| 111 |
+
- robust cross-lingual alignment,
|
| 112 |
+
- stable decoding for long outputs,
|
| 113 |
+
- strong zero-shot performance on French.
|
| 114 |
|
| 115 |
+
It is compatible with:
|
|
|
|
|
|
|
|
|
|
|
|
|
| 116 |
|
| 117 |
+
- Hugging Face Transformers,
|
| 118 |
+
- CTranslate2,
|
| 119 |
+
- ONNX Runtime,
|
| 120 |
+
- MLX (Apple Silicon),
|
| 121 |
+
- quantization-based acceleration pipelines.
|
| 122 |
|
| 123 |
---
|
| 124 |
|
| 125 |
+
## 5. Methodology and Reproducibility
|
| 126 |
|
| 127 |
+
`Gilbert-FR-Source` is used in strict research settings emphasizing:
|
| 128 |
|
| 129 |
+
### **Reproducible training protocols**
|
| 130 |
+
- frozen weights for baseline comparison,
|
| 131 |
+
- controlled hyperparameter schedules,
|
| 132 |
+
- consistent evaluation datasets,
|
| 133 |
+
- deterministic decoding configurations.
|
| 134 |
|
| 135 |
+
### **Evaluation methodology**
|
| 136 |
+
WER is computed with standard normalization (lowercasing, punctuation removal).
|
| 137 |
+
More advanced metrics (diarization error rate, long-context drift) are included in internal research pipelines.
|
| 138 |
+
|
| 139 |
+
### **Versioning policy**
|
| 140 |
+
This repository represents version `0.1` of the research baseline.
|
| 141 |
+
All future fine-tuned models will explicitly reference this version for traceability.
|
| 142 |
|
| 143 |
---
|
| 144 |
|
| 145 |
+
## 6. Limitations
|
| 146 |
+
|
| 147 |
+
This baseline inherits the known limitations of Whisper and of the underlying datasets:
|
| 148 |
|
| 149 |
+
- sensitivity to overlapping speech,
|
| 150 |
+
- occasional hallucinations in long-form decoding,
|
| 151 |
+
- domain shift on spontaneous dialogue,
|
| 152 |
+
- potential biases related to accent distribution in training data,
|
| 153 |
+
- suboptimal performance in telephony bandwidth.
|
| 154 |
+
|
| 155 |
+
Understanding and quantifying these limitations is one of the core objectives of the Gilbert research roadmap.
|
| 156 |
|
| 157 |
---
|
| 158 |
|
| 159 |
+
## 7. Future Work (Planned Research Directions)
|
| 160 |
+
|
| 161 |
+
The following models will be developed as independent checkpoints:
|
| 162 |
+
|
| 163 |
+
- **Gilbert-FR-Longform-v1**
|
| 164 |
+
Long meetings, multi-speaker interaction, discourse-level context stability.
|
| 165 |
|
| 166 |
+
- **Gilbert-FR-Accents-v1**
|
| 167 |
+
Robustness to regional and international French accents.
|
| 168 |
|
| 169 |
+
- **Gilbert-FR-Telephone-v1**
|
| 170 |
+
Optimized for 8 kHz VoIP/call-center speech.
|
| 171 |
|
| 172 |
+
- **Gilbert-Multilingual-v1**
|
| 173 |
+
Extended cross-lingual performance with optimized French anchors.
|
| 174 |
+
|
| 175 |
+
Each model will include detailed evaluation reports and will adhere to research reproducibility standards.
|
| 176 |
|
| 177 |
---
|
| 178 |
|
| 179 |
+
## 8. License
|
| 180 |
+
|
| 181 |
+
This repository includes files distributed under the MIT License.
|
| 182 |
|
| 183 |
+
> A copy of the MIT License is included.
|
| 184 |
+
> Some files were originally released under MIT.
|
|
|
|
|
|
|
| 185 |
|
| 186 |
+
All future Gilbert models built on top of this baseline are the exclusive property of Lexia France.
|
| 187 |
|
| 188 |
---
|
| 189 |
|
| 190 |
+
## 9. Contact
|
| 191 |
|
| 192 |
+
For research collaboration, evaluation access, or technical inquiries:
|
|
|
|
|
|
|
| 193 |
|
| 194 |
+
- Website: https://gilbert-assistant.fr
|
| 195 |
+
- Email: mathis@lexiapro.fr
|