auto-transcribe on stop_recording: submitting mid-recording silently sent nothing

app.py

@@ -15,6 +15,22 @@ MODEL = EncDecHybridRNNTCTCBPEModel.restore_from(_path, map_location="cpu")
 MODEL.eval()
 print("model preloaded", flush=True)
 
+DESCRIPTION = (
+    "121M Conformer trained on ~1,655 h of conversational Nepali. "
+    "**33.8% WER on real call-center audio** (NepTel benchmark) where Whisper-large-v3 "
+    "zero-shot scores ~99%. Honest limitations and the full benchmark: "
+    "[github.com/Ampixa/nepaliconformer](https://github.com/Ampixa/nepaliconformer). "
+    "CPU demo — a 30 s clip takes roughly 10-20 s. Example clips are real call-center "
+    "audio (CC-BY-4.0, © InfoBayAI)."
+)
+
+ARTICLE = (
+    "Recording transcribes itself as soon as you press ⏹ stop. "
+    "Mic blocked? [Open the demo full-screen](https://voidash-nepaliconformer.hf.space) · "
+    "Model downloads and usage: "
+    "[github.com/Ampixa/nepaliconformer](https://github.com/Ampixa/nepaliconformer#download--run)"
+)
+
 
 def get_model():
     return MODEL
@@ -59,7 +75,9 @@ def _prepare(path):
 
 def transcribe(audio_path):
     if not audio_path:
-        return "(no audio)"
+        # Most often: the user pressed Transcribe while the mic was still recording,
+        # so no file exists yet.
+        return "Press ⏹ stop to finish the recording — it transcribes automatically."
     audio_path, duration = _prepare(audio_path)
     if duration is not None and duration > 60:
         return "Please keep clips under 60 seconds for this CPU demo."
@@ -69,28 +87,37 @@ def transcribe(audio_path):
     return " ".join(t for t in text.split() if t != "<breath>") or "(no speech detected)"
 
 
-examples = [[f"examples/{f}"] for f in sorted(os.listdir("examples"))] \
-    if os.path.isdir("examples") else None
-
-demo = gr.Interface(
-    fn=transcribe,
-    inputs=gr.Audio(sources=["microphone", "upload"], type="filepath",
-                    label="Nepali speech (mic or file, ≤60 s)"),
-    outputs=gr.Textbox(label="Transcript (Devanagari)", rtl=False),
-    title="NepaliConformer — Nepali ASR for real telephone calls",
-    description=(
-        "121M Conformer trained on ~1,655 h of conversational Nepali. "
-        "**33.8% WER on real call-center audio** (NepTel benchmark) where Whisper-large-v3 "
-        "zero-shot scores ~99%. Honest limitations and the full benchmark: "
-        "[github.com/Ampixa/nepaliconformer](https://github.com/Ampixa/nepaliconformer). "
-        "CPU demo — a 30 s clip takes roughly 10-20 s. Example clips are real call-center "
-        "audio (CC-BY-4.0, © InfoBayAI)."
-    ),
-    examples=examples,
-    cache_examples=False,
-    flagging_mode="never",
-    article="Demo stuck or mic blocked? [Open the demo full-screen](https://voidash-nepaliconformer.hf.space) \u00b7 Model downloads and usage: [github.com/Ampixa/nepaliconformer](https://github.com/Ampixa/nepaliconformer#download--run)",
-)
+example_files = [[f"examples/{f}"] for f in sorted(os.listdir("examples"))] \
+    if os.path.isdir("examples") else []
+
+with gr.Blocks(title="NepaliConformer — Nepali ASR for real telephone calls") as demo:
+    gr.Markdown("# NepaliConformer — Nepali ASR for real telephone calls")
+    gr.Markdown(DESCRIPTION)
+    with gr.Row():
+        with gr.Column():
+            audio_in = gr.Audio(
+                sources=["microphone", "upload"], type="filepath",
+                label="Nepali speech (mic or file, ≤60 s)",
+            )
+            with gr.Row():
+                clear_btn = gr.Button("Clear")
+                submit_btn = gr.Button("Transcribe", variant="primary")
+        with gr.Column():
+            text_out = gr.Textbox(label="Transcript (Devanagari)", lines=6)
+
+    # Transcribe as soon as the recording stops or a file lands: waiting for an explicit
+    # Transcribe click made the demo look broken for anyone who never pressed stop.
+    audio_in.stop_recording(transcribe, audio_in, text_out)
+    audio_in.upload(transcribe, audio_in, text_out)
+    submit_btn.click(transcribe, audio_in, text_out)
+    clear_btn.click(lambda: (None, ""), None, [audio_in, text_out])
+
+    if example_files:
+        gr.Examples(
+            examples=example_files, inputs=audio_in, outputs=text_out,
+            fn=transcribe, cache_examples=False,
+        )
+    gr.Markdown(ARTICLE)
 
 if __name__ == "__main__":
     demo.launch(ssr_mode=False, show_error=True)