Update app.py

app.py

@@ -8,6 +8,21 @@ from backend.helpers import get_random_session_samples
 
 dataset = load_dataset("iyosha-huji/stressBench", token=configs.HF_API_TOKEN)["test"]
 
+import os
+import tempfile
+import soundfile as sf
+
+TEMP_AUDIO_DIR = os.path.join(tempfile.gettempdir(), "stress_eval_audio")
+os.makedirs(TEMP_AUDIO_DIR, exist_ok=True)
+
+def write_temp_wav(audio_array, sampling_rate, interpretation_id):
+    filename = f"{interpretation_id}.wav"
+    filepath = os.path.join(TEMP_AUDIO_DIR, filename)
+
+    if not os.path.exists(filepath):
+        sf.write(filepath, audio_array, sampling_rate)
+
+    return filepath
 
 def human_eval_tab():
     with gr.Tab(label="Evaluation"):
@@ -85,7 +100,7 @@ def human_eval_tab():
                         question_md = gr.Markdown()
                         radio = gr.Radio(label="Answer:", interactive=True)
                     with gr.Column(variant="compact"):
-                        audio_output = gr.Audio(interactive=False, type="numpy", label="Audio:")
+                        audio_output = gr.Audio(interactive=False, type="filepath", label="Audio:")
 
         with gr.Group(
             visible=False, elem_id="final_page"
@@ -116,13 +131,17 @@ def human_eval_tab():
             # show the question
             true_index = session_sample_indices[i]
             sample = dataset[true_index]
-            audio_data = (sample["audio"]["sampling_rate"], sample["audio"]["array"])
+            # audio_data = (sample["audio"]["sampling_rate"], sample["audio"]["array"])
+            # 🛠️ Save temp wav by interpretation_id
+            interp_id = sample["interpretation_id"]
+            wav_path = write_temp_wav(sample["audio"]["array"], sample["audio"]["sampling_rate"], interp_id)
+
             previous_answer = answers.get(i, None)
             return (
                 gr.update(visible=True),
                 f"<div align='center'>Sample <b>{i+1}</b> out of <b>{len(session_sample_indices)}</b></div>",
                 "Out of the following answers, according to the speaker's stressed words, what is most likely the underlying intention of the speaker?",
-                gr.update(value=audio_data),
+                gr.update(value=wav_path),
                 gr.update(
                     choices=sample["possible_answers"],
                     value=previous_answer,