Update app.py

app.py

@@ -15,15 +15,8 @@ model = AutoModelForSpeechSeq2Seq.from_pretrained(
     model_name, device_map=device, torch_dtype=torch.bfloat16
 )
 
-today_str = date.today().strftime("%B %d, %Y")
 
-system_prompt = (
-    "Knowledge Cutoff Date: April 2024.\n"
-    f"Today's Date: {today_str}.\n"
-    "You are Granite, developed by IBM. You are a helpful AI assistant."
-)
-
-def transcribe(audio_file):
+def transcribe(audio_file, user_prompt):
     # load wav file
     wav, sr = torchaudio.load(audio_file, normalize=True)
     if wav.shape[0] != 1 or sr != 16000:
@@ -32,20 +25,31 @@ def transcribe(audio_file):
         wav = torchaudio.functional.resample(wav, sr, 16000)
         sr = 16000
 
-    # user prompt
-    user_prompt = "<|audio|>can you transcribe the speech into a written format?"
+    today_str = date.today().strftime("%B %d, %Y")
+
+    system_prompt = (
+        "Knowledge Cutoff Date: April 2024.\n"
+        f"Today's Date: {today_str}.\n"
+        "You are Granite, developed by IBM. You are a helpful AI assistant."
+    )
+
     chat = [
         dict(role="system", content=system_prompt),
-        dict(role="user", content=user_prompt),
+        dict(role="user", content=f"<|audio|>{user_prompt}"),
     ]
-    prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
+    prompt = tokenizer.apply_chat_template(
+        chat, tokenize=False, add_generation_prompt=True)
 
     # run model
-    model_inputs = processor(prompt, wav, device=device, return_tensors="pt").to(device)
+    model_inputs = processor(
+        prompt,
+        wav,
+        device=device,
+        return_tensors="pt").to(device)
     model_outputs = model.generate(
-        **model_inputs, 
-        max_new_tokens=200, 
-        do_sample=False, 
+        **model_inputs,
+        max_new_tokens=512,
+        do_sample=False,
         num_beams=1
     )
 
@@ -58,15 +62,32 @@ def transcribe(audio_file):
 
     return output_text[0].strip()
 
+
 # Gradio UI
 with gr.Blocks() as demo:
-    gr.Markdown("## Granite 3.3 Speech-to-Text Demo")
+    gr.Markdown("## Granite 3.3 Speech-to-Text")
+    gr.Markdown(
+        "Upload an audio file and Granite Speech 3.3 8b will transcribe it into text."
+        "You can also edit the prompt below to customize what Granite should do with the audio, like translation."
+    )
 
     with gr.Row():
-        audio_input = gr.Audio(type="filepath", label="Upload Audio (16kHz mono preferred)")
+        audio_input = gr.Audio(type="filepath",
+                               label="Upload Audio (16kHz mono preferred)")
         output_text = gr.Textbox(label="Transcription", lines=5)
 
+    user_prompt = gr.Textbox(
+        label="User Prompt",
+        value="Can you transcribe the speech into a written format?",
+        lines=2
+    )
+
     transcribe_btn = gr.Button("Transcribe")
-    transcribe_btn.click(fn=transcribe, inputs=audio_input, outputs=output_text)
+    transcribe_btn.click(
+        fn=transcribe,
+        inputs=[
+            audio_input,
+            user_prompt],
+        outputs=output_text)
 
-    demo.launch()
+demo.launch()