up

app.py

@@ -55,6 +55,7 @@ max_range = np.iinfo(target_dtype).max
 
 
 def get_visualisation(idx):
+    idx -= 1
     audio = dataset[idx]["audio"]
     array = (audio["array"] * max_range).astype(np.int16)
     sampling_rate = audio["sampling_rate"]
@@ -64,33 +65,34 @@ def get_visualisation(idx):
 
     wer_output = process_words(text1, text2, wer_default, wer_default)
     wer_percentage = 100 * wer_output.wer
-    num_insertions = wer_output.insertions
+    rel_insertions = wer_output.insertions / len(text1.split())
 
     rel_length = len(text2.split()) / len(text1.split())
 
     diff = compare_string(text1, text2)
     full_text = style_text(diff)
 
-    return (sampling_rate, array), wer_percentage, num_insertions, rel_length, full_text
+    return (sampling_rate, array), wer_percentage, rel_insertions, rel_length, full_text
 
 
 if __name__ == "__main__":
+    gr.Markdown("Analyse the transcriptions generated by the Whisper large-v2 model on the TEDLIUM dev set.")
     with gr.Blocks() as demo:
         slider = gr.Slider(
-            minimum=0, maximum=len(norm_target), step=1, label="Dataset sample"
+            minimum=1, maximum=len(norm_target), step=1, label="Dataset sample"
         )
         btn = gr.Button("Analyse")
         audio_out = gr.Audio(label="Audio input")
         with gr.Row():
             wer = gr.Number(label="WER")
-            insertions = gr.Number(label="Insertions")
-            relative_length = gr.Number(label="Relative length of target / reference")
+            relative_insertions = gr.Number(label="Relative insertions (# insertions / target length)")
+            relative_length = gr.Number(label="Relative length (reference length / target length)")
         text_out = gr.Markdown(label="Text difference")
 
         btn.click(
             fn=get_visualisation,
             inputs=slider,
-            outputs=[audio_out, wer, insertions, relative_length, text_out],
+            outputs=[audio_out, wer, relative_insertions, relative_length, text_out],
         )
 
     demo.launch()