Update app.py

app.py

@@ -1,10 +1,12 @@
 import sys
+import re
 
 from importlib.metadata import version
 
 import evaluate
 import polars as pl
 import gradio as gr
+from natsort import natsorted
 
 # Load evaluators
 wer = evaluate.load("wer")
@@ -59,17 +61,45 @@ tech_env = f"""
 tech_libraries = f"""
 #### Libraries
 
-- evaluate: {version('evaluate')}
-- gradio: {version('gradio')}
-- jiwer: {version('jiwer')}
-- polars: {version('polars')}
+- evaluate: {version("evaluate")}
+- gradio: {version("gradio")}
+- jiwer: {version("jiwer")}
+- polars: {version("polars")}
 """.strip()
 
 
 def clean_value(x):
-    return x.replace('’', "'").strip().lower().replace(',', '').replace('.', '').replace('?', '').replace('!', '').replace('–', '').replace('«', '').replace('»', '')
+    s = (
+        x.replace("’", "'")
+        .strip()
+        .lower()
+        .replace(":", " ")
+        .replace(",", " ")
+        .replace(".", " ")
+        .replace("?", " ")
+        .replace("!", " ")
+        .replace("–", " ")
+        .replace("«", " ")
+        .replace("»", " ")
+        .replace("—", " ")
+        .replace("…", " ")
+        .replace("/", " ")
+        .replace("\\", " ")
+        .replace("(", " ")
+        .replace(")", " ")
+        .replace("́", "")
+        .replace('"', " ")
+    )
+
+    s = re.sub(r" +", " ", s)
+
+    s = s.strip()
+
+    # print(s)
+
+    return s
+
 
-    
 def inference(file_name, clear_punctuation, show_chars, progress=gr.Progress()):
     if not file_name:
         raise gr.Error("Please paste your JSON file.")
@@ -78,25 +108,21 @@ def inference(file_name, clear_punctuation, show_chars, progress=gr.Progress()):
 
     df = pl.read_ndjson(file_name)
 
-    inference_seconds = df['inference_total'].sum()
-    duration_seconds = df['duration'].sum()
+    inference_seconds = df["inference_total"].sum()
+    duration_seconds = df["duration"].sum()
 
     rtf = inference_seconds / duration_seconds
 
-    references = df['reference']
+    references = df["reference"]
 
     if clear_punctuation:
-        predictions = df['prediction'].map_elements(clean_value)
+        predictions = df["prediction"].map_elements(clean_value, return_dtype=pl.String)
     else:
-        predictions = df['prediction']
+        predictions = df["prediction"]
 
     # Evaluate
-    wer_value = round(
-        wer.compute(predictions=predictions, references=references), 4
-    )
-    cer_value = round(
-        cer.compute(predictions=predictions, references=references), 4
-    )
+    wer_value = round(wer.compute(predictions=predictions, references=references), 4)
+    cer_value = round(cer.compute(predictions=predictions, references=references), 4)
 
     inference_time = inference_seconds
     audio_duration = duration_seconds
@@ -106,27 +132,33 @@ def inference(file_name, clear_punctuation, show_chars, progress=gr.Progress()):
     results = []
 
     results.append(f"Metrics using `evaluate` library:")
-    results.append('')
-    results.append(f"- WER: {wer_value} metric, {round(wer_value*100, 4)}%")
-    results.append(f"- CER: {cer_value} metric, {round(cer_value*100, 4)}%")
-    results.append('')
+    results.append("")
+    results.append(f"- WER: {wer_value} metric, {round(wer_value * 100, 4)}%")
+    results.append(f"- CER: {cer_value} metric, {round(cer_value * 100, 4)}%")
+    results.append("")
     results.append(f"- Accuracy on words: {round(100 - 100 * wer_value, 4)}%")
     results.append(f"- Accuracy on chars: {round(100 - 100 * cer_value, 4)}%")
-    results.append('')
-    results.append(f"- Inference time: {round(inference_time, 4)} seconds, {round(inference_time/60, 4)} mins, {round(inference_time/60/60, 4)} hours")
-    results.append(f"- Audio duration: {round(audio_duration, 4)} seconds, {round(audio_duration/60/60, 4)} hours")
-    results.append('')
+    results.append("")
+    results.append(
+        f"- Inference time: {round(inference_time, 4)} seconds, {round(inference_time / 60, 4)} mins, {round(inference_time / 60 / 60, 4)} hours"
+    )
+    results.append(
+        f"- Audio duration: {round(audio_duration, 4)} seconds, {round(audio_duration / 60 / 60, 4)} hours"
+    )
+    results.append("")
     results.append(f"- RTF: {round(rtf, 4)}")
 
     if show_chars:
         all_chars = set()
-        for pred in list(df['prediction']):
+        for pred in predictions:
             for c in pred:
                 all_chars.add(c)
 
-        results.append('')
+        sorted_chars = natsorted(list(all_chars))
+
+        results.append("")
         results.append(f"Chars in predictions:")
-        results.append(f"{list(all_chars)}")
+        results.append(f"{sorted_chars}")
 
     return "\n".join(results)
 
@@ -161,12 +193,16 @@ with demo:
     gr.Button("Calculate").click(
         inference,
         concurrency_limit=concurrency_limit,
-        inputs=[jsonl_file, clear_punctuation],
+        inputs=[jsonl_file, clear_punctuation, show_chars],
         outputs=metrics,
     )
 
     with gr.Row():
-        gr.Examples(label="Choose an example", inputs=[jsonl_file, clear_punctuation, show_chars], examples=examples)
+        gr.Examples(
+            label="Choose an example",
+            inputs=[jsonl_file, clear_punctuation, show_chars],
+            examples=examples,
+        )
 
     gr.Markdown(description_foot)