add wer

README.md

@@ -45,7 +45,7 @@ model-index:
         metrics:
           - name: Test WER
             type: wer
-            value: ??
+            value: 7.88
       - task:
           type: Automatic Speech Recognition
           name: automatic-speech-recognition
@@ -59,7 +59,7 @@ model-index:
         metrics:
           - name: Test WER
             type: wer
-            value: ??
+            value: 4.57
       - task:
           type: Automatic Speech Recognition
           name: automatic-speech-recognition
@@ -73,7 +73,7 @@ model-index:
         metrics:
           - name: Test WER
             type: wer
-            value: ??
+            value: 10.14
       - task:
           type: Automatic Speech Recognition
           name: automatic-speech-recognition
@@ -87,7 +87,7 @@ model-index:
         metrics:
           - name: Test WER
             type: wer
-            value: ??
+            value: 19.8
 ---
 # LinTO STT French – FastConformer
 
@@ -109,7 +109,6 @@ This model is a fine-tuned version of the [NVIDIA French FastConformer Hybrid La
 It is a large (115M parameters) hybrid ASR model trained with both **Transducer (default)** and **CTC** losses.
 
 Compared to the base model, this version:
-- Does **not** include punctuation or uppercase letters.
 - Was trained on **10,000+ hours** of diverse, manually transcribed French speech.
 ---
 
@@ -130,6 +129,10 @@ Evaluations can be very long (especially for whisper) so we selected only segmen
 - Multilingual TEDx: 972 rows (1.5h)
 - 50% of our internal Youtube corpus: 956 rows (1h)
 
+![WER table](https://huggingface.co/linagora/linto_stt_fr_fastconformer_pc/resolve/main/assets/wer_table.png)
+
+As shown in the table above (lower is better), the model demonstrates robust performance across all datasets, consistently achieving results close to the best even on the out-of-domain dataset like SUMM-RE.
+
 ---
 
 ## Usage
@@ -188,7 +191,7 @@ or be deployed within a microservices infrastructure using a message broker conn
 ### Data Setup
 - Audios ranging from 0.1s to 30s
 - Batch size of 50
-- All datasets except YODAS and TouTubeFR were upsampled
+- All datasets except YODAS and TouTubeFR were upsampled 2x
 
 ### Training datasets