Check point 4

app.py

@@ -10,15 +10,17 @@ import torchaudio
 from scipy.spatial.distance import cosine
 from RealtimeSTT import AudioToTextRecorder
 from fastapi import FastAPI, APIRouter
-from fastrtc import Stream, AsyncStreamHandler, ReplyOnPause, get_cloudflare_turn_credentials_async, get_cloudflare_turn_credentials
+from fastrtc import Stream, AsyncStreamHandler
 import json
-import io
-import wave
 import asyncio
 import uvicorn
-import socket
 from queue import Queue
-import time
+import logging
+
+# Set up logging
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
 # Simplified configuration parameters
 SILENCE_THRESHS = [0, 0.4]
 FINAL_TRANSCRIPTION_MODEL = "distil-large-v3"
@@ -32,35 +34,31 @@ MIN_LENGTH_OF_RECORDING = 0.7
 PRE_RECORDING_BUFFER_DURATION = 0.35
 
 # Speaker change detection parameters
-DEFAULT_CHANGE_THRESHOLD = 0.7
+DEFAULT_CHANGE_THRESHOLD = 0.65
 EMBEDDING_HISTORY_SIZE = 5
-MIN_SEGMENT_DURATION = 1.0
+MIN_SEGMENT_DURATION = 1.5
 DEFAULT_MAX_SPEAKERS = 4
-ABSOLUTE_MAX_SPEAKERS = 10
+ABSOLUTE_MAX_SPEAKERS = 8
 
 # Global variables
-FAST_SENTENCE_END = True
 SAMPLE_RATE = 16000
-BUFFER_SIZE = 512
+BUFFER_SIZE = 1024
 CHANNELS = 1
 
-# Speaker colors
+# Speaker colors - more distinguishable colors
 SPEAKER_COLORS = [
-    "#FFFF00",  # Yellow
-    "#FF0000",  # Red
-    "#00FF00",  # Green
-    "#00FFFF",  # Cyan
-    "#FF00FF",  # Magenta
-    "#0000FF",  # Blue
-    "#FF8000",  # Orange
-    "#00FF80",  # Spring Green
-    "#8000FF",  # Purple
-    "#FFFFFF",  # White
+    "#FF6B6B",  # Red
+    "#4ECDC4",  # Teal
+    "#45B7D1",  # Blue
+    "#96CEB4",  # Green
+    "#FFEAA7",  # Yellow
+    "#DDA0DD",  # Plum
+    "#98D8C8",  # Mint
+    "#F7DC6F",  # Gold
 ]
 
 SPEAKER_COLOR_NAMES = [
-    "Yellow", "Red", "Green", "Cyan", "Magenta", 
-    "Blue", "Orange", "Spring Green", "Purple", "White"
+    "Red", "Teal", "Blue", "Green", "Yellow", "Plum", "Mint", "Gold"
 ]
 
 
@@ -74,24 +72,11 @@ class SpeechBrainEncoder:
         self.cache_dir = os.path.join(os.path.expanduser("~"), ".cache", "speechbrain")
         os.makedirs(self.cache_dir, exist_ok=True)
     
-    def _download_model(self):
-        """Download pre-trained SpeechBrain ECAPA-TDNN model if not present"""
-        model_url = "https://huggingface.co/speechbrain/spkrec-ecapa-voxceleb/resolve/main/embedding_model.ckpt"
-        model_path = os.path.join(self.cache_dir, "embedding_model.ckpt")
-        
-        if not os.path.exists(model_path):
-            print(f"Downloading ECAPA-TDNN model to {model_path}...")
-            urllib.request.urlretrieve(model_url, model_path)
-        
-        return model_path
-    
     def load_model(self):
         """Load the ECAPA-TDNN model"""
         try:
             from speechbrain.pretrained import EncoderClassifier
             
-            model_path = self._download_model()
-            
             self.model = EncoderClassifier.from_hparams(
                 source="speechbrain/spkrec-ecapa-voxceleb",
                 savedir=self.cache_dir,
@@ -99,9 +84,10 @@ class SpeechBrainEncoder:
             )
             
             self.model_loaded = True
+            logger.info("ECAPA-TDNN model loaded successfully!")
             return True
         except Exception as e:
-            print(f"Error loading ECAPA-TDNN model: {e}")
+            logger.error(f"Error loading ECAPA-TDNN model: {e}")
             return False
     
     def embed_utterance(self, audio, sr=16000):
@@ -111,10 +97,15 @@ class SpeechBrainEncoder:
         
         try:
             if isinstance(audio, np.ndarray):
-                waveform = torch.tensor(audio, dtype=torch.float32).unsqueeze(0)
+                # Ensure audio is float32 and properly normalized
+                audio = audio.astype(np.float32)
+                if np.max(np.abs(audio)) > 1.0:
+                    audio = audio / np.max(np.abs(audio))
+                waveform = torch.tensor(audio).unsqueeze(0)
             else:
                 waveform = audio.unsqueeze(0)
             
+            # Resample if necessary
             if sr != 16000:
                 waveform = torchaudio.functional.resample(waveform, orig_freq=sr, new_freq=16000)
             
@@ -123,7 +114,7 @@ class SpeechBrainEncoder:
                 
             return embedding.squeeze().cpu().numpy()
         except Exception as e:
-            print(f"Error extracting embedding: {e}")
+            logger.error(f"Error extracting embedding: {e}")
             return np.zeros(self.embedding_dim)
 
 
@@ -131,41 +122,60 @@ class AudioProcessor:
     """Processes audio data to extract speaker embeddings"""
     def __init__(self, encoder):
         self.encoder = encoder
+        self.audio_buffer = []
+        self.min_audio_length = int(SAMPLE_RATE * 1.0)  # Minimum 1 second of audio
     
-    def extract_embedding(self, audio_int16):
-        try:
-            float_audio = audio_int16.astype(np.float32) / 32768.0
+    def add_audio_chunk(self, audio_chunk):
+        """Add audio chunk to buffer"""
+        self.audio_buffer.extend(audio_chunk)
+        
+        # Keep buffer from getting too large
+        max_buffer_size = int(SAMPLE_RATE * 10)  # 10 seconds max
+        if len(self.audio_buffer) > max_buffer_size:
+            self.audio_buffer = self.audio_buffer[-max_buffer_size:]
+    
+    def extract_embedding_from_buffer(self):
+        """Extract embedding from current audio buffer"""
+        if len(self.audio_buffer) < self.min_audio_length:
+            return None
             
-            if np.abs(float_audio).max() > 1.0:
-                float_audio = float_audio / np.abs(float_audio).max()
+        try:
+            # Use the last portion of the buffer for embedding
+            audio_segment = np.array(self.audio_buffer[-self.min_audio_length:], dtype=np.float32)
             
-            embedding = self.encoder.embed_utterance(float_audio)
+            # Normalize audio
+            if np.max(np.abs(audio_segment)) > 0:
+                audio_segment = audio_segment / np.max(np.abs(audio_segment))
+            else:
+                return None
             
+            embedding = self.encoder.embed_utterance(audio_segment)
             return embedding
         except Exception as e:
-            print(f"Embedding extraction error: {e}")
-            return np.zeros(self.encoder.embedding_dim)
+            logger.error(f"Embedding extraction error: {e}")
+            return None
 
 
 class SpeakerChangeDetector:
-    """Speaker change detector that supports a configurable number of speakers"""
+    """Improved speaker change detector"""
     def __init__(self, embedding_dim=192, change_threshold=DEFAULT_CHANGE_THRESHOLD, max_speakers=DEFAULT_MAX_SPEAKERS):
         self.embedding_dim = embedding_dim
         self.change_threshold = change_threshold
         self.max_speakers = min(max_speakers, ABSOLUTE_MAX_SPEAKERS)
         self.current_speaker = 0
-        self.previous_embeddings = []
-        self.last_change_time = time.time()
-        self.mean_embeddings = [None] * self.max_speakers
         self.speaker_embeddings = [[] for _ in range(self.max_speakers)]
-        self.last_similarity = 0.0
+        self.speaker_centroids = [None] * self.max_speakers
+        self.last_change_time = time.time()
+        self.last_similarity = 1.0
         self.active_speakers = set([0])
+        self.segment_counter = 0
         
     def set_max_speakers(self, max_speakers):
         """Update the maximum number of speakers"""
         new_max = min(max_speakers, ABSOLUTE_MAX_SPEAKERS)
         
         if new_max < self.max_speakers:
+            # Remove speakers beyond the new limit
             for speaker_id in list(self.active_speakers):
                 if speaker_id >= new_max:
                     self.active_speakers.discard(speaker_id)
@@ -173,85 +183,85 @@ class SpeakerChangeDetector:
             if self.current_speaker >= new_max:
                 self.current_speaker = 0
         
+        # Resize arrays
         if new_max > self.max_speakers:
-            self.mean_embeddings.extend([None] * (new_max - self.max_speakers))
             self.speaker_embeddings.extend([[] for _ in range(new_max - self.max_speakers)])
+            self.speaker_centroids.extend([None] * (new_max - self.max_speakers))
         else:
-            self.mean_embeddings = self.mean_embeddings[:new_max]
             self.speaker_embeddings = self.speaker_embeddings[:new_max]
+            self.speaker_centroids = self.speaker_centroids[:new_max]
         
         self.max_speakers = new_max
         
     def set_change_threshold(self, threshold):
         """Update the threshold for detecting speaker changes"""
-        self.change_threshold = max(0.1, min(threshold, 0.99))
+        self.change_threshold = max(0.1, min(threshold, 0.95))
         
     def add_embedding(self, embedding, timestamp=None):
-        """Add a new embedding and check if there's a speaker change"""
+        """Add a new embedding and detect speaker changes"""
         current_time = timestamp or time.time()
-        
-        if not self.previous_embeddings:
-            self.previous_embeddings.append(embedding)
-            self.speaker_embeddings[self.current_speaker].append(embedding)
-            if self.mean_embeddings[self.current_speaker] is None:
-                self.mean_embeddings[self.current_speaker] = embedding.copy()
-            return self.current_speaker, 1.0
-        
-        current_mean = self.mean_embeddings[self.current_speaker]
-        if current_mean is not None:
-            similarity = 1.0 - cosine(embedding, current_mean)
+        self.segment_counter += 1
+        
+        # Initialize first speaker
+        if not self.speaker_embeddings[0]:
+            self.speaker_embeddings[0].append(embedding)
+            self.speaker_centroids[0] = embedding.copy()
+            self.active_speakers.add(0)
+            return 0, 1.0
+        
+        # Calculate similarity with current speaker
+        current_centroid = self.speaker_centroids[self.current_speaker]
+        if current_centroid is not None:
+            similarity = 1.0 - cosine(embedding, current_centroid)
         else:
-            similarity = 1.0 - cosine(embedding, self.previous_embeddings[-1])
+            similarity = 0.5
         
         self.last_similarity = similarity
         
+        # Check for speaker change
         time_since_last_change = current_time - self.last_change_time
-        is_speaker_change = False
+        speaker_changed = False
         
-        if time_since_last_change >= MIN_SEGMENT_DURATION:
-            if similarity < self.change_threshold:
-                best_speaker = self.current_speaker
-                best_similarity = similarity
-                
-                for speaker_id in range(self.max_speakers):
-                    if speaker_id == self.current_speaker:
-                        continue
-                        
-                    speaker_mean = self.mean_embeddings[speaker_id]
+        if time_since_last_change >= MIN_SEGMENT_DURATION and similarity < self.change_threshold:
+            # Find best matching speaker
+            best_speaker = self.current_speaker
+            best_similarity = similarity
+            
+            for speaker_id in self.active_speakers:
+                if speaker_id == self.current_speaker:
+                    continue
                     
-                    if speaker_mean is not None:
-                        speaker_similarity = 1.0 - cosine(embedding, speaker_mean)
-                        
-                        if speaker_similarity > best_similarity:
-                            best_similarity = speaker_similarity
-                            best_speaker = speaker_id
-                
-                if best_speaker != self.current_speaker:
-                    is_speaker_change = True
-                    self.current_speaker = best_speaker
-                elif len(self.active_speakers) < self.max_speakers:
-                    for new_id in range(self.max_speakers):
-                        if new_id not in self.active_speakers:
-                            is_speaker_change = True
-                            self.current_speaker = new_id
-                            self.active_speakers.add(new_id)
-                            break
-        
-        if is_speaker_change:
-            self.last_change_time = current_time
-        
-        self.previous_embeddings.append(embedding)
-        if len(self.previous_embeddings) > EMBEDDING_HISTORY_SIZE:
-            self.previous_embeddings.pop(0)
-        
+                centroid = self.speaker_centroids[speaker_id]
+                if centroid is not None:
+                    speaker_similarity = 1.0 - cosine(embedding, centroid)
+                    if speaker_similarity > best_similarity and speaker_similarity > self.change_threshold:
+                        best_similarity = speaker_similarity
+                        best_speaker = speaker_id
+            
+            # If no good match found and we can add a new speaker
+            if best_speaker == self.current_speaker and len(self.active_speakers) < self.max_speakers:
+                for new_id in range(self.max_speakers):
+                    if new_id not in self.active_speakers:
+                        best_speaker = new_id
+                        self.active_speakers.add(new_id)
+                        break
+            
+            if best_speaker != self.current_speaker:
+                self.current_speaker = best_speaker
+                self.last_change_time = current_time
+                speaker_changed = True
+        
+        # Update speaker embeddings and centroids
         self.speaker_embeddings[self.current_speaker].append(embedding)
-        self.active_speakers.add(self.current_speaker)
         
-        if len(self.speaker_embeddings[self.current_speaker]) > 30:
-            self.speaker_embeddings[self.current_speaker] = self.speaker_embeddings[self.current_speaker][-30:]
-            
+        # Keep only recent embeddings (sliding window)
+        max_embeddings = 20
+        if len(self.speaker_embeddings[self.current_speaker]) > max_embeddings:
+            self.speaker_embeddings[self.current_speaker] = self.speaker_embeddings[self.current_speaker][-max_embeddings:]
+        
+        # Update centroid
         if self.speaker_embeddings[self.current_speaker]:
-            self.mean_embeddings[self.current_speaker] = np.mean(
+            self.speaker_centroids[self.current_speaker] = np.mean(
                 self.speaker_embeddings[self.current_speaker], axis=0
             )
         
@@ -264,7 +274,7 @@ class SpeakerChangeDetector:
         return "#FFFFFF"
     
     def get_status_info(self):
-        """Return status information about the speaker change detector"""
+        """Return status information"""
         speaker_counts = [len(self.speaker_embeddings[i]) for i in range(self.max_speakers)]
         
         return {
@@ -273,7 +283,8 @@ class SpeakerChangeDetector:
             "active_speakers": len(self.active_speakers),
             "max_speakers": self.max_speakers,
             "last_similarity": self.last_similarity,
-            "threshold": self.change_threshold
+            "threshold": self.change_threshold,
+            "segment_counter": self.segment_counter
         }
 
 
@@ -287,19 +298,18 @@ class RealtimeSpeakerDiarization:
         self.full_sentences = []
         self.sentence_speakers = []
         self.pending_sentences = []
-        self.displayed_text = ""
-        self.last_realtime_text = ""
+        self.current_conversation = ""
         self.is_running = False
         self.change_threshold = DEFAULT_CHANGE_THRESHOLD
         self.max_speakers = DEFAULT_MAX_SPEAKERS
-        self.current_conversation = ""
-        self.audio_buffer = []
+        self.last_transcription = ""
+        self.transcription_lock = threading.Lock()
         
     def initialize_models(self):
         """Initialize the speaker encoder model"""
         try:
             device_str = "cuda" if torch.cuda.is_available() else "cpu"
-            print(f"Using device: {device_str}")
+            logger.info(f"Using device: {device_str}")
             
             self.encoder = SpeechBrainEncoder(device=device_str)
             success = self.encoder.load_model()
@@ -311,80 +321,94 @@ class RealtimeSpeakerDiarization:
                     change_threshold=self.change_threshold,
                     max_speakers=self.max_speakers
                 )
-                print("ECAPA-TDNN model loaded successfully!")
+                logger.info("Models initialized successfully!")
                 return True
             else:
-                print("Failed to load ECAPA-TDNN model")
+                logger.error("Failed to load models")
                 return False
         except Exception as e:
-            print(f"Model initialization error: {e}")
+            logger.error(f"Model initialization error: {e}")
             return False
     
     def live_text_detected(self, text):
         """Callback for real-time transcription updates"""
-        text = text.strip()
-        if text:
-            sentence_delimiters = '.?!。'
-            prob_sentence_end = (
-                len(self.last_realtime_text) > 0
-                and text[-1] in sentence_delimiters
-                and self.last_realtime_text[-1] in sentence_delimiters
-            )
-
-            self.last_realtime_text = text
-
-            if prob_sentence_end and FAST_SENTENCE_END:
-                self.recorder.stop()
-            elif prob_sentence_end:
-                self.recorder.post_speech_silence_duration = SILENCE_THRESHS[0]
-            else:
-                self.recorder.post_speech_silence_duration = SILENCE_THRESHS[1]
+        with self.transcription_lock:
+            self.last_transcription = text.strip()
     
     def process_final_text(self, text):
         """Process final transcribed text with speaker embedding"""
         text = text.strip()
         if text:
             try:
-                bytes_data = self.recorder.last_transcription_bytes
-                self.sentence_queue.put((text, bytes_data))
-                self.pending_sentences.append(text)
+                # Get audio data for this transcription
+                audio_bytes = getattr(self.recorder, 'last_transcription_bytes', None)
+                if audio_bytes:
+                    self.sentence_queue.put((text, audio_bytes))
+                else:
+                    # If no audio bytes, use current speaker
+                    self.sentence_queue.put((text, None))
+                    
             except Exception as e:
-                print(f"Error processing final text: {e}")
+                logger.error(f"Error processing final text: {e}")
     
     def process_sentence_queue(self):
         """Process sentences in the queue for speaker detection"""
         while self.is_running:
             try:
-                text, bytes_data = self.sentence_queue.get(timeout=1)
-                
-                # Convert audio data to int16
-                audio_int16 = np.frombuffer(bytes_data, dtype=np.int16)
-                
-                # Extract speaker embedding
-                speaker_embedding = self.audio_processor.extract_embedding(audio_int16)
+                text, audio_bytes = self.sentence_queue.get(timeout=1)
                 
-                # Store sentence and embedding
-                self.full_sentences.append((text, speaker_embedding))
+                current_speaker = self.speaker_detector.current_speaker
                 
-                # Fill in missing speaker assignments
-                while len(self.sentence_speakers) < len(self.full_sentences) - 1:
-                    self.sentence_speakers.append(0)
-                
-                # Detect speaker changes
-                speaker_id, similarity = self.speaker_detector.add_embedding(speaker_embedding)
-                self.sentence_speakers.append(speaker_id)
-                
-                # Remove from pending
-                if text in self.pending_sentences:
-                    self.pending_sentences.remove(text)
+                if audio_bytes:
+                    # Convert audio data and extract embedding
+                    audio_int16 = np.frombuffer(audio_bytes, dtype=np.int16)
+                    audio_float = audio_int16.astype(np.float32) / 32768.0
+                    
+                    # Extract embedding
+                    embedding = self.audio_processor.encoder.embed_utterance(audio_float)
+                    if embedding is not None:
+                        current_speaker, similarity = self.speaker_detector.add_embedding(embedding)
                 
-                # Update conversation display
-                self.current_conversation = self.get_formatted_conversation()
+                # Store sentence with speaker
+                with self.transcription_lock:
+                    self.full_sentences.append((text, current_speaker))
+                    self.update_conversation_display()
                     
             except queue.Empty:
                 continue
             except Exception as e:
-                print(f"Error processing sentence: {e}")
+                logger.error(f"Error processing sentence: {e}")
+    
+    def update_conversation_display(self):
+        """Update the conversation display"""
+        try:
+            sentences_with_style = []
+            
+            for sentence_text, speaker_id in self.full_sentences:
+                color = self.speaker_detector.get_color_for_speaker(speaker_id)
+                speaker_name = f"Speaker {speaker_id + 1}"
+                sentences_with_style.append(
+                    f'<span style="color:{color}; font-weight: bold;">{speaker_name}:</span> '
+                    f'<span style="color:#333333;">{sentence_text}</span>'
+                )
+            
+            # Add current transcription if available
+            if self.last_transcription:
+                current_color = self.speaker_detector.get_color_for_speaker(self.speaker_detector.current_speaker)
+                current_speaker = f"Speaker {self.speaker_detector.current_speaker + 1}"
+                sentences_with_style.append(
+                    f'<span style="color:{current_color}; font-weight: bold; opacity: 0.7;">{current_speaker}:</span> '
+                    f'<span style="color:#666666; font-style: italic;">{self.last_transcription}...</span>'
+                )
+            
+            if sentences_with_style:
+                self.current_conversation = "<br><br>".join(sentences_with_style)
+            else:
+                self.current_conversation = "<i>Waiting for speech input...</i>"
+                
+        except Exception as e:
+            logger.error(f"Error updating conversation display: {e}")
+            self.current_conversation = f"<i>Error: {str(e)}</i>"
     
     def start_recording(self):
         """Start the recording and transcription process"""
@@ -392,10 +416,10 @@ class RealtimeSpeakerDiarization:
             return "Please initialize models first!"
         
         try:
-            # Setup recorder configuration for manual audio input
+            # Setup recorder configuration
             recorder_config = {
                 'spinner': False,
-                'use_microphone': False,  # We'll feed audio manually
+                'use_microphone': True,  # Changed to True for direct microphone input
                 'model': FINAL_TRANSCRIPTION_MODEL,
                 'language': TRANSCRIPTION_LANGUAGE,
                 'silero_sensitivity': SILERO_SENSITIVITY,
@@ -405,29 +429,28 @@ class RealtimeSpeakerDiarization:
                 'pre_recording_buffer_duration': PRE_RECORDING_BUFFER_DURATION,
                 'min_gap_between_recordings': 0,
                 'enable_realtime_transcription': True,
-                'realtime_processing_pause': 0,
+                'realtime_processing_pause': 0.1,
                 'realtime_model_type': REALTIME_TRANSCRIPTION_MODEL,
                 'on_realtime_transcription_update': self.live_text_detected,
                 'beam_size': FINAL_BEAM_SIZE,
                 'beam_size_realtime': REALTIME_BEAM_SIZE,
-                'buffer_size': BUFFER_SIZE,
                 'sample_rate': SAMPLE_RATE,
             }
 
             self.recorder = AudioToTextRecorder(**recorder_config)
             
-            # Start sentence processing thread
+            # Start processing threads
             self.is_running = True
             self.sentence_thread = threading.Thread(target=self.process_sentence_queue, daemon=True)
             self.sentence_thread.start()
             
-            # Start transcription thread
             self.transcription_thread = threading.Thread(target=self.run_transcription, daemon=True)
             self.transcription_thread.start()
             
-            return "Recording started successfully! FastRTC audio input ready."
+            return "Recording started successfully!"
             
         except Exception as e:
+            logger.error(f"Error starting recording: {e}")
             return f"Error starting recording: {e}"
     
     def run_transcription(self):
@@ -436,7 +459,7 @@ class RealtimeSpeakerDiarization:
             while self.is_running:
                 self.recorder.text(self.process_final_text)
         except Exception as e:
-            print(f"Transcription error: {e}")
+            logger.error(f"Transcription error: {e}")
     
     def stop_recording(self):
         """Stop the recording process"""
@@ -447,12 +470,10 @@ class RealtimeSpeakerDiarization:
     
     def clear_conversation(self):
         """Clear all conversation data"""
-        self.full_sentences = []
-        self.sentence_speakers = []
-        self.pending_sentences = []
-        self.displayed_text = ""
-        self.last_realtime_text = ""
-        self.current_conversation = "Conversation cleared!"
+        with self.transcription_lock:
+            self.full_sentences = []
+            self.last_transcription = ""
+            self.current_conversation = "Conversation cleared!"
         
         if self.speaker_detector:
             self.speaker_detector = SpeakerChangeDetector(
@@ -475,36 +496,8 @@ class RealtimeSpeakerDiarization:
         return f"Settings updated: Threshold={threshold:.2f}, Max Speakers={max_speakers}"
     
     def get_formatted_conversation(self):
-        """Get the formatted conversation with speaker colors"""
-        try:
-            sentences_with_style = []
-            
-            # Process completed sentences
-            for i, sentence in enumerate(self.full_sentences):
-                sentence_text, _ = sentence
-                if i >= len(self.sentence_speakers):
-                    color = "#FFFFFF"
-                    speaker_name = "Unknown"
-                else:
-                    speaker_id = self.sentence_speakers[i]
-                    color = self.speaker_detector.get_color_for_speaker(speaker_id)
-                    speaker_name = f"Speaker {speaker_id + 1}"
-                    
-                sentences_with_style.append(
-                    f'<span style="color:{color};"><b>{speaker_name}:</b> {sentence_text}</span>')
-            
-            # Add pending sentences
-            for pending_sentence in self.pending_sentences:
-                sentences_with_style.append(
-                    f'<span style="color:#60FFFF;"><b>Processing:</b> {pending_sentence}</span>')
-            
-            if sentences_with_style:
-                return "<br><br>".join(sentences_with_style)
-            else:
-                return "Waiting for speech input..."
-                
-        except Exception as e:
-            return f"Error formatting conversation: {e}"
+        """Get the formatted conversation"""
+        return self.current_conversation
     
     def get_status_info(self):
         """Get current status information"""
@@ -520,689 +513,427 @@ class RealtimeSpeakerDiarization:
                 f"**Last Similarity:** {status['last_similarity']:.3f}",
                 f"**Change Threshold:** {status['threshold']:.2f}",
                 f"**Total Sentences:** {len(self.full_sentences)}",
+                f"**Segments Processed:** {status['segment_counter']}",
                 "",
-                "**Speaker Segment Counts:**"
+                "**Speaker Activity:**"
             ]
             
             for i in range(status['max_speakers']):
                 color_name = SPEAKER_COLOR_NAMES[i] if i < len(SPEAKER_COLOR_NAMES) else f"Speaker {i+1}"
-                status_lines.append(f"Speaker {i+1} ({color_name}): {status['speaker_counts'][i]}")
+                count = status['speaker_counts'][i]
+                active = "🟢" if count > 0 else "⚫"
+                status_lines.append(f"{active} Speaker {i+1} ({color_name}): {count} segments")
             
             return "\n".join(status_lines)
             
         except Exception as e:
             return f"Error getting status: {e}"
 
-    def feed_audio_data(self, audio_data):
-        """Feed audio data to the recorder"""
-        if not self.is_running or not self.recorder:
-            return
-        
-        try:
-            # Ensure audio is in the correct format (16-bit PCM)
-            if isinstance(audio_data, np.ndarray):
-                if audio_data.dtype != np.int16:
-                    # Convert float to int16
-                    if audio_data.dtype == np.float32 or audio_data.dtype == np.float64:
-                        audio_data = (audio_data * 32767).astype(np.int16)
-                    else:
-                        audio_data = audio_data.astype(np.int16)
-                
-                # Convert to bytes
-                audio_bytes = audio_data.tobytes()
-            else:
-                audio_bytes = audio_data
-            
-            # Feed to recorder
-            self.recorder.feed_audio(audio_bytes)
-            
-        except Exception as e:
-            print(f"Error feeding audio data: {e}")
-
     def process_audio_chunk(self, audio_data, sample_rate=16000):
         """Process audio chunk from FastRTC input"""
-        if not self.is_running or self.recorder is None:
+        if not self.is_running or self.audio_processor is None:
             return
             
         try:
-            # Convert float audio to int16 for the recorder
-            if audio_data.dtype == np.float32 or audio_data.dtype == np.float64:
-                if np.max(np.abs(audio_data)) <= 1.0:
-                    # Float audio is normalized to [-1, 1], convert to int16
-                    audio_int16 = (audio_data * 32767).astype(np.int16)
-                else:
-                    # Audio is already in higher range
-                    audio_int16 = audio_data.astype(np.int16)
+            # Ensure audio is float32
+            if isinstance(audio_data, np.ndarray):
+                if audio_data.dtype != np.float32:
+                    audio_data = audio_data.astype(np.float32)
             else:
-                audio_int16 = audio_data
-                
-            # Ensure correct shape (1, N) for the recorder
-            if len(audio_int16.shape) == 1:
-                audio_int16 = np.expand_dims(audio_int16, 0)
-                
-            # Resample if needed
-            if sample_rate != SAMPLE_RATE:
-                audio_int16 = self._resample_audio(audio_int16, sample_rate, SAMPLE_RATE)
-                
-            # Convert to bytes for feeding to recorder
-            audio_bytes = audio_int16.tobytes()
-            
-            # Feed to recorder
-            self.feed_audio_data(audio_bytes)
+                audio_data = np.array(audio_data, dtype=np.float32)
             
-        except Exception as e:
-            print(f"Error processing audio chunk: {e}")
-            
-    def _resample_audio(self, audio, orig_sr, target_sr):
-        """Resample audio to target sample rate"""
-        try:
-            import scipy.signal
+            # Ensure mono
+            if len(audio_data.shape) > 1:
+                audio_data = np.mean(audio_data, axis=1) if audio_data.shape[1] > 1 else audio_data.flatten()
             
-            # Get the resampling ratio
-            ratio = target_sr / orig_sr
+            # Normalize if needed
+            if np.max(np.abs(audio_data)) > 1.0:
+                audio_data = audio_data / np.max(np.abs(audio_data))
             
-            # Calculate the new length
-            new_length = int(len(audio[0]) * ratio)
+            # Add to audio processor buffer for speaker detection
+            self.audio_processor.add_audio_chunk(audio_data)
             
-            # Resample the audio
-            resampled = scipy.signal.resample(audio[0], new_length)
-            
-            # Return in the same shape format
-            return np.expand_dims(resampled, 0)
+            # Periodically extract embeddings for speaker detection
+            if len(self.audio_processor.audio_buffer) % (SAMPLE_RATE // 2) == 0:  # Every 0.5 seconds
+                embedding = self.audio_processor.extract_embedding_from_buffer()
+                if embedding is not None:
+                    self.speaker_detector.add_embedding(embedding)
+                    
         except Exception as e:
-            print(f"Error resampling audio: {e}")
-            return audio
-
+            logger.error(f"Error processing audio chunk: {e}")
 
-# FastRTC Audio Handler for Real-time Diarization
 
+# FastRTC Audio Handler
 class DiarizationHandler(AsyncStreamHandler):
     def __init__(self, diarization_system):
         super().__init__()
         self.diarization_system = diarization_system
-        self.audio_queue = Queue()
-        self.is_processing = False
-        self.sample_rate = 16000  # Default sample rate
+        self.audio_buffer = []
+        self.buffer_size = BUFFER_SIZE
         
     def copy(self):
         """Return a fresh handler for each new stream connection"""
         return DiarizationHandler(self.diarization_system)
     
     async def emit(self):
-        """Not used in this implementation - we only receive audio"""
+        """Not used - we only receive audio"""
         return None
     
     async def receive(self, frame):
-        """Receive audio data from FastRTC and process it"""
+        """Receive audio data from FastRTC"""
         try:
             if not self.diarization_system.is_running:
                 return
                 
-            # Extract audio data from frame
-            if hasattr(frame, 'data') and frame.data is not None:
-                audio_data = frame.data
-            elif hasattr(frame, 'audio') and frame.audio is not None:
-                audio_data = frame.audio
-            else:
-                audio_data = frame
+            # Extract audio data
+            audio_data = getattr(frame, 'data', frame)
             
-            # Convert to numpy array if needed
+            # Convert to numpy array
             if isinstance(audio_data, bytes):
-                # Convert bytes to numpy array (assuming 16-bit PCM)
-                audio_array = np.frombuffer(audio_data, dtype=np.int16)
-                # Normalize to float32 range [-1, 1]
-                audio_array = audio_array.astype(np.float32) / 32768.0
+                audio_array = np.frombuffer(audio_data, dtype=np.int16).astype(np.float32) / 32768.0
             elif isinstance(audio_data, (list, tuple)):
                 audio_array = np.array(audio_data, dtype=np.float32)
-            elif isinstance(audio_data, np.ndarray):
-                audio_array = audio_data.astype(np.float32)
             else:
-                print(f"Unknown audio data type: {type(audio_data)}")
-                return
-            
-            # Ensure mono audio
-            if len(audio_array.shape) > 1 and audio_array.shape[1] > 1:
-                audio_array = np.mean(audio_array, axis=1)
+                audio_array = np.array(audio_data, dtype=np.float32)
             
-            # Ensure 1D array
+            # Ensure 1D
             if len(audio_array.shape) > 1:
                 audio_array = audio_array.flatten()
             
-            # Get sample rate from frame if available
-            sample_rate = getattr(frame, 'sample_rate', self.sample_rate)
+            # Buffer audio chunks
+            self.audio_buffer.extend(audio_array)
             
-            # Process audio asynchronously to avoid blocking
-            await self.process_audio_async(audio_array, sample_rate)
+            # Process in chunks
+            while len(self.audio_buffer) >= self.buffer_size:
+                chunk = np.array(self.audio_buffer[:self.buffer_size])
+                self.audio_buffer = self.audio_buffer[self.buffer_size:]
+                
+                # Process asynchronously
+                await self.process_audio_async(chunk)
                 
         except Exception as e:
-            print(f"Error in FastRTC audio receive: {e}")
-            import traceback
-            traceback.print_exc()
+            logger.error(f"Error in FastRTC receive: {e}")
     
-    async def process_audio_async(self, audio_data, sample_rate=16000):
+    async def process_audio_async(self, audio_data):
         """Process audio data asynchronously"""
         try:
-            # Run the audio processing in a thread pool to avoid blocking
             loop = asyncio.get_event_loop()
             await loop.run_in_executor(
                 None, 
                 self.diarization_system.process_audio_chunk, 
                 audio_data, 
-                sample_rate
+                SAMPLE_RATE
             )
         except Exception as e:
-            print(f"Error in async audio processing: {e}")
-            
-    async def start_up(self) -> None:
-        """Initialize any resources when the stream starts"""
-        print("FastRTC stream started")
-        self.is_processing = True
-        
-    async def shutdown(self) -> None:
-        """Clean up any resources when the stream ends"""
-        print("FastRTC stream shutting down")
-        self.is_processing = False
+            logger.error(f"Error in async audio processing: {e}")
 
 
 # Global instances
-diarization_system = None  # Will be initialized when RealtimeSpeakerDiarization is available
+diarization_system = RealtimeSpeakerDiarization()
 audio_handler = None
 
-
 def initialize_system():
     """Initialize the diarization system"""
-    global audio_handler, diarization_system
+    global audio_handler
     try:
-        if diarization_system is None:
-            print("Error: RealtimeSpeakerDiarization not initialized")
-            return "❌ Diarization system not available. Please ensure RealtimeSpeakerDiarization is properly imported."
-        
         success = diarization_system.initialize_models()
         if success:
             audio_handler = DiarizationHandler(diarization_system)
-            return "✅ System initialized successfully! Models loaded and FastRTC handler ready."
+            return "✅ System initialized successfully!"
         else:
-            return "❌ Failed to initialize system. Please check the logs."
+            return "❌ Failed to initialize system. Check logs for details."
     except Exception as e:
-        print(f"Initialization error: {e}")
+        logger.error(f"Initialization error: {e}")
         return f"❌ Initialization error: {str(e)}"
 
-
 def start_recording():
     """Start recording and transcription"""
     try:
-        if diarization_system is None:
-            return "❌ System not initialized"
         result = diarization_system.start_recording()
-        return f"🎙️ {result} - FastRTC audio streaming is active."
+        return f"🎙️ {result}"
     except Exception as e:
         return f"❌ Failed to start recording: {str(e)}"
 
-
 def stop_recording():
     """Stop recording and transcription"""
     try:
-        if diarization_system is None:
-            return "❌ System not initialized"
         result = diarization_system.stop_recording()
         return f"⏹️ {result}"
     except Exception as e:
         return f"❌ Failed to stop recording: {str(e)}"
 
-
 def clear_conversation():
     """Clear the conversation"""
     try:
-        if diarization_system is None:
-            return "❌ System not initialized"
         result = diarization_system.clear_conversation()
         return f"🗑️ {result}"
     except Exception as e:
         return f"❌ Failed to clear conversation: {str(e)}"
 
-
 def update_settings(threshold, max_speakers):
     """Update system settings"""
     try:
-        if diarization_system is None:
-            return "❌ System not initialized"
         result = diarization_system.update_settings(threshold, max_speakers)
         return f"⚙️ {result}"
     except Exception as e:
         return f"❌ Failed to update settings: {str(e)}"
 
-
 def get_conversation():
     """Get the current conversation"""
     try:
-        if diarization_system is None:
-            return "<i>System not initialized</i>"
         return diarization_system.get_formatted_conversation()
     except Exception as e:
         return f"<i>Error getting conversation: {str(e)}</i>"
 
-
 def get_status():
     """Get system status"""
     try:
-        if diarization_system is None:
-            return "System not initialized"
         return diarization_system.get_status_info()
     except Exception as e:
         return f"Error getting status: {str(e)}"
 
-
 # Create Gradio interface
 def create_interface():
     with gr.Blocks(title="Real-time Speaker Diarization", theme=gr.themes.Soft()) as interface:
         gr.Markdown("# 🎤 Real-time Speech Recognition with Speaker Diarization")
-        gr.Markdown("This app performs real-time speech recognition with automatic speaker identification using FastRTC for low-latency audio streaming.")
+        gr.Markdown("Live transcription with automatic speaker identification using FastRTC audio streaming.")
         
         with gr.Row():
             with gr.Column(scale=2):
-                # Main conversation display
+                # Conversation display
                 conversation_output = gr.HTML(
-                    value="<div style='padding: 20px; background: #f5f5f5; border-radius: 10px;'><i>Click 'Initialize System' to start...</i></div>",
-                    label="Live Conversation",
-                    elem_id="conversation_display"
+                    value="<div style='padding: 20px; background: #f8f9fa; border-radius: 10px; min-height: 300px;'><i>Click 'Initialize System' to start...</i></div>",
+                    label="Live Conversation"
                 )
                 
                 # Control buttons
                 with gr.Row():
                     init_btn = gr.Button("🔧 Initialize System", variant="secondary", size="lg")
-                    start_btn = gr.Button("🎙️ Start Recording", variant="primary", size="lg", interactive=False)
-                    stop_btn = gr.Button("⏹️ Stop Recording", variant="stop", size="lg", interactive=False)
+                    start_btn = gr.Button("🎙️ Start", variant="primary", size="lg", interactive=False)
+                    stop_btn = gr.Button("⏹️ Stop", variant="stop", size="lg", interactive=False)
                     clear_btn = gr.Button("🗑️ Clear", variant="secondary", size="lg", interactive=False)
                 
-                # FastRTC Stream Interface
-                with gr.Row():
-                    gr.HTML("""
-                    <div id="fastrtc-container" style="border: 2px solid #ddd; border-radius: 10px; padding: 20px; margin: 10px 0;">
-                        <h3>🎵 Audio Stream</h3>
-                        <p>FastRTC audio stream will appear here when recording starts.</p>
-                        <div id="stream-status" style="padding: 10px; background: #f8f9fa; border-radius: 5px; margin-top: 10px;">
-                            Status: Waiting for initialization...
-                        </div>
-                    </div>
-                    """)
-                
                 # Status display
                 status_output = gr.Textbox(
                     label="System Status",
-                    value="System not initialized. Please click 'Initialize System' to begin.",
-                    lines=6,
-                    interactive=False,
-                    show_copy_button=True
+                    value="Ready to initialize...",
+                    lines=8,
+                    interactive=False
                 )
             
             with gr.Column(scale=1):
-                # Settings panel
+                # Settings
                 gr.Markdown("## ⚙️ Settings")
                 
                 threshold_slider = gr.Slider(
-                    minimum=0.1,
-                    maximum=0.95,
+                    minimum=0.3,
+                    maximum=0.9,
                     step=0.05,
-                    value=0.5,  # DEFAULT_CHANGE_THRESHOLD
+                    value=DEFAULT_CHANGE_THRESHOLD,
                     label="Speaker Change Sensitivity",
-                    info="Lower = more sensitive to speaker changes"
+                    info="Lower = more sensitive"
                 )
                 
                 max_speakers_slider = gr.Slider(
                     minimum=2,
-                    maximum=10,  # ABSOLUTE_MAX_SPEAKERS
+                    maximum=ABSOLUTE_MAX_SPEAKERS,
                     step=1,
-                    value=4,  # DEFAULT_MAX_SPEAKERS
-                    label="Maximum Number of Speakers"
+                    value=DEFAULT_MAX_SPEAKERS,
+                    label="Maximum Speakers"
                 )
                 
-                update_settings_btn = gr.Button("Update Settings", variant="secondary")
-                
-                # Audio settings
-                gr.Markdown("## 🔊 Audio Configuration")
-                with gr.Accordion("Advanced Audio Settings", open=False):
-                    gr.Markdown("""
-                    **Current Configuration:**
-                    - Sample Rate: 16kHz
-                    - Audio Format: 16-bit PCM → Float32 (via AudioProcessor)
-                    - Channels: Mono (stereo converted automatically)
-                    - Buffer Size: 1024 samples for real-time processing
-                    - Processing: Uses existing AudioProcessor.extract_embedding()
-                    """)
+                update_btn = gr.Button("Update Settings", variant="secondary")
                 
                 # Instructions
-                gr.Markdown("## 📝 How to Use")
                 gr.Markdown("""
-                1. **Initialize**: Click "Initialize System" to load AI models
-                2. **Start**: Click "Start Recording" to begin processing  
-                3. **Connect**: The FastRTC stream will activate automatically
-                4. **Allow Access**: Grant microphone permissions when prompted
-                5. **Speak**: Talk naturally into your microphone
-                6. **Monitor**: Watch real-time transcription with speaker colors
-                """)
-                
-                # Performance tips
-                with gr.Accordion("💡 Performance Tips", open=False):
-                    gr.Markdown("""
-                    - Use Chrome/Edge for best FastRTC performance
-                    - Ensure stable internet connection
-                    - Use headphones to prevent echo
-                    - Position microphone 6-12 inches away
-                    - Minimize background noise
-                    - Allow browser microphone access
-                    """)
+                ## 📋 Instructions
+                1. **Initialize** the system (loads AI models)
+                2. **Start** recording 
+                3. **Speak** - system will transcribe and identify speakers
+                4. **Monitor** real-time results below
                 
-                # Speaker color legend
-                gr.Markdown("## 🎨 Speaker Colors")
-                speaker_colors = [
-                    ("#FF6B6B", "Red"),
-                    ("#4ECDC4", "Teal"), 
-                    ("#45B7D1", "Blue"),
-                    ("#96CEB4", "Green"),
-                    ("#FFEAA7", "Yellow"),
-                    ("#DDA0DD", "Plum"),
-                    ("#98D8C8", "Mint"),
-                    ("#F7DC6F", "Gold")
-                ]
-                
-                color_html = ""
-                for i, (color, name) in enumerate(speaker_colors[:4]):
-                    color_html += f'<div style="margin: 3px 0;"><span style="color:{color}; font-size: 16px; font-weight: bold;">●</span> Speaker {i+1} ({name})</div>'
-                
-                gr.HTML(f"<div style='font-size: 14px;'>{color_html}</div>")
-        
-        # Auto-refresh conversation and status
-        def refresh_display():
-            try:
-                conversation = get_conversation()
-                status = get_status()
-                return conversation, status
-            except Exception as e:
-                error_msg = f"Error refreshing display: {str(e)}"
-                return f"<i>{error_msg}</i>", error_msg
+                ## 🎨 Speaker Colors
+                - 🔴 Speaker 1 (Red)
+                - 🟢 Speaker 2 (Teal) 
+                - 🔵 Speaker 3 (Blue)
+                - 🟡 Speaker 4 (Green)
+                - 🟣 Speaker 5 (Yellow)
+                - 🟤 Speaker 6 (Plum)
+                - 🟫 Speaker 7 (Mint)
+                - 🟨 Speaker 8 (Gold)
+                """)
         
         # Event handlers
         def on_initialize():
-            try:
-                result = initialize_system()
-                success = "successfully" in result.lower()
-                
-                conversation, status = refresh_display()
-                
-                return (
-                    result,  # status_output
-                    gr.update(interactive=success),   # start_btn
-                    gr.update(interactive=success),   # clear_btn
-                    conversation,  # conversation_output
-                )
-            except Exception as e:
-                error_msg = f"❌ Initialization failed: {str(e)}"
-                return (
-                    error_msg,
-                    gr.update(interactive=False),
-                    gr.update(interactive=False),
-                    "<i>System not ready</i>",
-                )
+            result = initialize_system()
+            if "✅" in result:
+                return result, gr.update(interactive=True), gr.update(interactive=True), gr.update(interactive=True)
+            else:
+                return result, gr.update(interactive=False), gr.update(interactive=False), gr.update(interactive=False)
         
         def on_start():
-            try:
-                result = start_recording()
-                return (
-                    result,  # status_output
-                    gr.update(interactive=False),  # start_btn
-                    gr.update(interactive=True),   # stop_btn
-                )
-            except Exception as e:
-                error_msg = f"❌ Failed to start: {str(e)}"
-                return (
-                    error_msg,
-                    gr.update(interactive=True),
-                    gr.update(interactive=False),
-                )
+            result = start_recording()
+            return result, gr.update(interactive=False), gr.update(interactive=True)
         
         def on_stop():
-            try:
-                result = stop_recording()
-                return (
-                    result,  # status_output
-                    gr.update(interactive=True),   # start_btn
-                    gr.update(interactive=False),  # stop_btn
-                )
-            except Exception as e:
-                error_msg = f"❌ Failed to stop: {str(e)}"
-                return (
-                    error_msg,
-                    gr.update(interactive=False),
-                    gr.update(interactive=True),
-                )
+            result = stop_recording()
+            return result, gr.update(interactive=True), gr.update(interactive=False)
         
         def on_clear():
-            try:
-                result = clear_conversation()
-                conversation, status = refresh_display()
-                return result, conversation
-            except Exception as e:
-                error_msg = f"❌ Failed to clear: {str(e)}"
-                return error_msg, "<i>Error clearing conversation</i>"
+            result = clear_conversation()
+            return result
         
         def on_update_settings(threshold, max_speakers):
-            try:
-                result = update_settings(threshold, max_speakers)
-                return result
-            except Exception as e:
-                return f"❌ Failed to update settings: {str(e)}"
+            result = update_settings(threshold, int(max_speakers))
+            return result
+        
+        def refresh_conversation():
+            return get_conversation()
         
-        # Connect event handlers
+        def refresh_status():
+            return get_status()
+        
+        # Button click handlers
         init_btn.click(
-            on_initialize,
-            outputs=[status_output, start_btn, clear_btn, conversation_output]
+            fn=on_initialize,
+            outputs=[status_output, start_btn, stop_btn, clear_btn]
         )
         
         start_btn.click(
-            on_start,
+            fn=on_start,
             outputs=[status_output, start_btn, stop_btn]
         )
         
         stop_btn.click(
-            on_stop,
+            fn=on_stop,
             outputs=[status_output, start_btn, stop_btn]
         )
         
         clear_btn.click(
-            on_clear,
-            outputs=[status_output, conversation_output]
+            fn=on_clear,
+            outputs=[status_output]
         )
         
-        update_settings_btn.click(
-            on_update_settings,
+        update_btn.click(
+            fn=on_update_settings,
             inputs=[threshold_slider, max_speakers_slider],
             outputs=[status_output]
         )
         
-        # Auto-refresh every 2 seconds when active
-        refresh_timer = gr.Timer(2.0)
-        refresh_timer.tick(
-            refresh_display,
-            outputs=[conversation_output, status_output]
-        )
+        # Create timers for auto-refreshing components
+        conversation_timer = gr.Timer(1, refresh_conversation, outputs=[conversation_output])
+        status_timer = gr.Timer(2, refresh_status, outputs=[status_output])
     
     return interface
 
 
-# FastAPI setup for API endpoints
-def create_fastapi_app():
-    """Create FastAPI app with API endpoints"""
-    app = FastAPI(
-        title="Real-time Speaker Diarization",
-        description="Real-time speech recognition with speaker diarization using FastRTC",
-        version="1.0.0"
-    )
-    
-    # API Routes
-    router = APIRouter()
-    
-    @router.get("/health")
-    async def health_check():
-        """Health check endpoint"""
-        return {
-            "status": "healthy",
-            "timestamp": time.time(),
-            "system_initialized": diarization_system is not None and hasattr(diarization_system, 'encoder') and diarization_system.encoder is not None,
-            "recording_active": diarization_system.is_running if diarization_system and hasattr(diarization_system, 'is_running') else False
-        }
-    
-    @router.get("/api/conversation")
-    async def get_conversation_api():
-        """Get current conversation"""
-        try:
-            return {
-                "conversation": get_conversation(),
-                "status": get_status(),
-                "is_recording": diarization_system.is_running if diarization_system and hasattr(diarization_system, 'is_running') else False,
-                "timestamp": time.time()
-            }
-        except Exception as e:
-            return {"error": str(e), "timestamp": time.time()}
-    
-    @router.post("/api/control/{action}")
-    async def control_recording(action: str):
-        """Control recording actions"""
-        try:
-            if action == "start":
-                result = start_recording()
-            elif action == "stop":
-                result = stop_recording()
-            elif action == "clear":
-                result = clear_conversation()
-            elif action == "initialize":
-                result = initialize_system()
-            else:
-                return {"error": "Invalid action. Use: start, stop, clear, or initialize"}
-            
-            return {
-                "result": result, 
-                "is_recording": diarization_system.is_running if diarization_system and hasattr(diarization_system, 'is_running') else False,
-                "timestamp": time.time()
-            }
-        except Exception as e:
-            return {"error": str(e), "timestamp": time.time()}
-    
-    app.include_router(router)
-    return app
+# FastAPI setup for FastRTC integration
+app = FastAPI()
 
+@app.get("/")
+async def root():
+    return {"message": "Real-time Speaker Diarization API"}
 
-# Function to setup FastRTC stream
-def setup_fastrtc_stream(app):
-    """Setup FastRTC stream with proper configuration"""
-    try:
-        if audio_handler is None:
-            print("Warning: Audio handler not initialized. Initialize system first.")
-            return None
-        
-        # Get HuggingFace token for TURN server (optional)
-        hf_token = os.environ.get("HF_TOKEN")
-        
-        # Configure RTC settings
-        rtc_config = {
-            "iceServers": [
-                {"urls": "stun:stun.l.google.com:19302"},
-                {"urls": "stun:stun1.l.google.com:19302"}
-            ]
-        }
-        
-        # Create FastRTC stream
-        stream = Stream(
-            handler=audio_handler,
-            rtc_configuration=rtc_config,
-            modality="audio",
-            mode="receive"  # We only receive audio, don't send
-        )
-        
-        # Mount the stream
-        app.mount("/stream", stream)
-        print("✅ FastRTC stream configured successfully!")
-        return stream
-        
-    except Exception as e:
-        print(f"⚠️ Warning: Failed to setup FastRTC stream: {e}")
-        print("Audio streaming may not work properly.")
-        return None
+@app.get("/health")
+async def health_check():
+    return {"status": "healthy", "system_running": diarization_system.is_running}
 
+@app.post("/initialize")
+async def api_initialize():
+    result = initialize_system()
+    return {"result": result, "success": "✅" in result}
 
-# Main application setup
-def create_app(diarization_sys=None):
-    """Create the complete application"""
-    global diarization_system
-    
-    # Set the diarization system
-    if diarization_sys is not None:
-        diarization_system = diarization_sys
-    
-    # Create FastAPI app
-    fastapi_app = create_fastapi_app()
-    
-    # Create Gradio interface
-    gradio_interface = create_interface()
-    
-    # Mount Gradio on FastAPI
-    app = gr.mount_gradio_app(fastapi_app, gradio_interface, path="/")
-    
-    # Setup FastRTC stream
-    if diarization_system is not None:
-        # Initialize the system if not already done
-        if not hasattr(diarization_system, 'encoder') or diarization_system.encoder is None:
-            diarization_system.initialize_models()
-        
-        # Create audio handler if needed
-        global audio_handler
-        if audio_handler is None:
-            audio_handler = DiarizationHandler(diarization_system)
-        
-        # Setup and mount the FastRTC stream
-        setup_fastrtc_stream(app)
-    
-    return app, gradio_interface
+@app.post("/start")
+async def api_start():
+    result = start_recording()
+    return {"result": result, "success": "🎙️" in result}
+
+@app.post("/stop")
+async def api_stop():
+    result = stop_recording()
+    return {"result": result, "success": "⏹️" in result}
+
+@app.post("/clear")
+async def api_clear():
+    result = clear_conversation()
+    return {"result": result}
+
+@app.get("/conversation")
+async def api_get_conversation():
+    return {"conversation": get_conversation()}
+
+@app.get("/status")
+async def api_get_status():
+    return {"status": get_status()}
+
+@app.post("/settings")
+async def api_update_settings(threshold: float, max_speakers: int):
+    result = update_settings(threshold, max_speakers)
+    return {"result": result}
+
+# FastRTC Stream setup
+if audio_handler:
+    stream = Stream(handler=audio_handler)
+    app.include_router(stream.router, prefix="/stream")
 
 
-# Entry point for HuggingFace Spaces
+# Main execution
 if __name__ == "__main__":
-    try:
-        # Import your diarization system here
-        # from your_module import RealtimeSpeakerDiarization
-        diarization_system = RealtimeSpeakerDiarization()
-        
-        # Create the application
-        app, interface = create_app()
-        
-        # Launch for HuggingFace Spaces
+    import argparse
+    
+    parser = argparse.ArgumentParser(description="Real-time Speaker Diarization System")
+    parser.add_argument("--mode", choices=["gradio", "api", "both"], default="gradio", 
+                       help="Run mode: gradio interface, API only, or both")
+    parser.add_argument("--host", default="0.0.0.0", help="Host to bind to")
+    parser.add_argument("--port", type=int, default=7860, help="Port to bind to")
+    parser.add_argument("--api-port", type=int, default=8000, help="API port (when running both)")
+    
+    args = parser.parse_args()
+    
+    if args.mode == "gradio":
+        # Run Gradio interface only
+        interface = create_interface()
         interface.launch(
-            server_name="0.0.0.0",
-            server_port=7860,
+            server_name=args.host,
+            server_port=args.port,
             share=True,
-            show_error=True,
-            quiet=False
+            show_error=True
         )
+    
+    elif args.mode == "api":
+        # Run FastAPI only
+        uvicorn.run(
+            app, 
+            host=args.host, 
+            port=args.port,
+            log_level="info"
+        )
+    
+    elif args.mode == "both":
+        # Run both Gradio and FastAPI
+        import multiprocessing
+        import threading
         
-    except Exception as e:
-        print(f"Failed to launch application: {e}")
-        import traceback
-        traceback.print_exc()
-        
-        # Fallback - launch just Gradio interface
-        try:
+        def run_gradio():
             interface = create_interface()
             interface.launch(
-                server_name="0.0.0.0",
-                server_port=int(os.environ.get("PORT", 7860)),
-                share=False
+                server_name=args.host,
+                server_port=args.port,
+                share=True,
+                show_error=True
             )
-        except Exception as fallback_error:
-            print(f"Fallback launch also failed: {fallback_error}")
-
-
-# Helper function to initialize with your diarization system
-def initialize_with_diarization_system(diarization_sys):
-    """Initialize the application with your diarization system"""
-    global diarization_system
-    diarization_system = diarization_sys
-    return create_app(diarization_sys)
+        
+        def run_fastapi():
+            uvicorn.run(
+                app,
+                host=args.host,
+                port=args.api_port,
+                log_level="info"
+            )
+        
+        # Start FastAPI in a separate thread
+        api_thread = threading.Thread(target=run_fastapi, daemon=True)
+        api_thread.start()
+        
+        # Start Gradio in main thread
+        run_gradio()