Update app.py

app.py

@@ -1,8 +1,9 @@
+import streamlit as st
 import torch
 import torchaudio
 import os
 import re
-import streamlit as st
+import jieba
 from difflib import SequenceMatcher
 from transformers import pipeline
 
@@ -16,46 +17,82 @@ pipe = pipeline(
     task="automatic-speech-recognition",
     model=MODEL_NAME,
     chunk_length_s=60,
-    device=device
+    device=device,
+    generate_kwargs={
+        "no_repeat_ngram_size": 4,
+        "repetition_penalty": 1.15,
+        "temperature": 0.5,
+        "top_p": 0.97,
+        "top_k": 40,
+        "max_new_tokens": 300,
+        "do_sample": True
+    }
 )
 pipe.model.config.forced_decoder_ids = pipe.tokenizer.get_decoder_prompt_ids(language=language, task="transcribe")
 
-# Load quality rating model
-rating_pipe = pipeline("text-classification", model="tabularisai/multilingual-sentiment-analysis")
+def is_similar(a, b, threshold=0.8):
+    return SequenceMatcher(None, a, b).ratio() > threshold
 
-# Sentiment label mapping
-label_map = {"Very Negative": "Very Poor", "Negative": "Poor", "Neutral": "Neutral", "Positive": "Good", "Very Positive": "Very Good"}
+def remove_repeated_phrases(text):
+    sentences = re.split(r'(?<=[。！？])', text)
+    cleaned_sentences = []
+    for i, sentence in enumerate(sentences):
+        if i == 0 or not is_similar(sentence.strip(), cleaned_sentences[-1].strip()):
+            cleaned_sentences.append(sentence.strip())
+    return " ".join(cleaned_sentences)
 
 def remove_punctuation(text):
     return re.sub(r'[^\w\s]', '', text)
 
 def transcribe_audio(audio_path):
-    transcript = pipe(audio_path)["text"]
-    return remove_punctuation(transcript)
+    waveform, sample_rate = torchaudio.load(audio_path)
+    duration = waveform.shape[1] / sample_rate
+    if duration > 60:
+        results = []
+        for start in range(0, int(duration), 55):
+            end = min(start + 60, int(duration))
+            chunk = waveform[:, start * sample_rate:end * sample_rate]
+            if chunk.shape[1] == 0:
+                continue
+            temp_filename = f"temp_chunk_{start}.wav"
+            torchaudio.save(temp_filename, chunk, sample_rate)
+            if os.path.exists(temp_filename):
+                try:
+                    result = pipe(temp_filename)["text"]
+                    results.append(remove_punctuation(result))
+                finally:
+                    os.remove(temp_filename)
+        return remove_punctuation(remove_repeated_phrases(" ".join(results)))
+    return remove_punctuation(remove_repeated_phrases(pipe(audio_path)["text"]))
+
+# Load quality rating model
+rating_pipe = pipeline("text-classification", model="tabularisai/multilingual-sentiment-analysis")
 
 def rate_quality(text):
-    result = rating_pipe(text)[0]
-    return label_map.get(result["label"], "Unknown")
+    chunks = [text[i:i+512] for i in range(0, len(text), 512)]
+    results = []
+    label_map = {"Very Negative": "Very Poor", "Negative": "Poor", "Neutral": "Neutral", "Positive": "Good", "Very Positive": "Very Good"}
+    
+    for chunk in chunks:
+        result = rating_pipe(chunk)[0]
+        results.append(label_map.get(result["label"], "Unknown"))
+    
+    return max(set(results), key=results.count)
 
 # Streamlit UI
-st.set_page_config(page_title="Cantonese Audio Transcription & Analysis", layout="centered")
-st.title("🗣️ Customer Service Conversation Quality Analyzer")
-st.markdown("Upload your Cantonese audio file, and we will transcribe and analyze its sentiment.")
-
-uploaded_file = st.file_uploader("Upload an audio file (WAV, MP3, etc.)", type=["wav", "mp3", "m4a"])
-if uploaded_file is not None:
-    with st.spinner("Processing audio..."):
-        temp_audio_path = "temp_audio.wav"
-        with open(temp_audio_path, "wb") as f:
-            f.write(uploaded_file.getbuffer())
-        transcript = transcribe_audio(temp_audio_path)
-        sentiment = rate_quality(transcript)
-        os.remove(temp_audio_path)
-    
-    st.subheader("Transcription")
-    st.text_area("", transcript, height=150)
+st.title("Cantonese Audio Transcription and Quality Rating")
+st.write("Upload your Cantonese audio file to get the transcription and quality rating.")
+
+audio_file = st.file_uploader("Upload Audio File", type=["wav", "mp3", "m4a"])
+
+if audio_file is not None:
+    audio_path = audio_file.name
+    with open(audio_path, "wb") as f:
+        f.write(audio_file.getbuffer())
     
-    st.subheader("Sentiment Analysis")
-    st.markdown(f"### 🎭 Sentiment: **{sentiment}**")
+    st.write("Processing audio...")
+    transcript = transcribe_audio(audio_path)
+    st.write("**Transcript:**", transcript)
     
-    st.success("Processing complete! 🎉")
+    quality_rating = rate_quality(transcript)
+    st.write("**Quality Rating:**", quality_rating)