app.py

import gradio as gr
import numpy as np
import queue
import torch
import time
import threading
import os
import urllib.request
import torchaudio
from scipy.spatial.distance import cosine
import json
import io
import wave

# Simplified configuration parameters
SILENCE_THRESHS = [0, 0.4]
FINAL_TRANSCRIPTION_MODEL = "distil-large-v3"
FINAL_BEAM_SIZE = 5
REALTIME_TRANSCRIPTION_MODEL = "distil-small.en"
REALTIME_BEAM_SIZE = 5
TRANSCRIPTION_LANGUAGE = "en"
SILERO_SENSITIVITY = 0.4
WEBRTC_SENSITIVITY = 3
MIN_LENGTH_OF_RECORDING = 0.7
PRE_RECORDING_BUFFER_DURATION = 0.35

# Speaker change detection parameters
DEFAULT_CHANGE_THRESHOLD = 0.7
EMBEDDING_HISTORY_SIZE = 5
MIN_SEGMENT_DURATION = 1.0
DEFAULT_MAX_SPEAKERS = 4
ABSOLUTE_MAX_SPEAKERS = 10

# Global variables
FAST_SENTENCE_END = True
SAMPLE_RATE = 16000
BUFFER_SIZE = 512
CHANNELS = 1

# Speaker colors
SPEAKER_COLORS = [
    "#FFFF00",  # Yellow
    "#FF0000",  # Red
    "#00FF00",  # Green
    "#00FFFF",  # Cyan
    "#FF00FF",  # Magenta
    "#0000FF",  # Blue
    "#FF8000",  # Orange
    "#00FF80",  # Spring Green
    "#8000FF",  # Purple
    "#FFFFFF",  # White
]

SPEAKER_COLOR_NAMES = [
    "Yellow", "Red", "Green", "Cyan", "Magenta", 
    "Blue", "Orange", "Spring Green", "Purple", "White"
]


class SpeechBrainEncoder:
    """ECAPA-TDNN encoder from SpeechBrain for speaker embeddings"""
    def __init__(self, device="cpu"):
        self.device = device
        self.model = None
        self.embedding_dim = 192
        self.model_loaded = False
        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,
                run_opts={"device": self.device}
            )
            
            self.model_loaded = True
            return True
        except Exception as e:
            print(f"Error loading ECAPA-TDNN model: {e}")
            return False
    
    def embed_utterance(self, audio, sr=16000):
        """Extract speaker embedding from audio"""
        if not self.model_loaded:
            raise ValueError("Model not loaded. Call load_model() first.")
        
        try:
            if isinstance(audio, np.ndarray):
                waveform = torch.tensor(audio, dtype=torch.float32).unsqueeze(0)
            else:
                waveform = audio.unsqueeze(0)
            
            if sr != 16000:
                waveform = torchaudio.functional.resample(waveform, orig_freq=sr, new_freq=16000)
            
            with torch.no_grad():
                embedding = self.model.encode_batch(waveform)
                
            return embedding.squeeze().cpu().numpy()
        except Exception as e:
            print(f"Error extracting embedding: {e}")
            return np.zeros(self.embedding_dim)


class AudioProcessor:
    """Processes audio data to extract speaker embeddings"""
    def __init__(self, encoder):
        self.encoder = encoder
    
    def extract_embedding(self, audio_float):
        try:
            # Ensure audio is in the right format
            if np.abs(audio_float).max() > 1.0:
                audio_float = audio_float / np.abs(audio_float).max()
            
            embedding = self.encoder.embed_utterance(audio_float)
            
            return embedding
        except Exception as e:
            print(f"Embedding extraction error: {e}")
            return np.zeros(self.encoder.embedding_dim)


class SpeakerChangeDetector:
    """Speaker change detector that supports a configurable number of speakers"""
    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.active_speakers = set([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:
            for speaker_id in list(self.active_speakers):
                if speaker_id >= new_max:
                    self.active_speakers.discard(speaker_id)
            
            if self.current_speaker >= new_max:
                self.current_speaker = 0
        
        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)])
        else:
            self.mean_embeddings = self.mean_embeddings[:new_max]
            self.speaker_embeddings = self.speaker_embeddings[: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))
        
    def add_embedding(self, embedding, timestamp=None):
        """Add a new embedding and check if there's a speaker change"""
        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)
        else:
            similarity = 1.0 - cosine(embedding, self.previous_embeddings[-1])
        
        self.last_similarity = similarity
        
        time_since_last_change = current_time - self.last_change_time
        is_speaker_change = 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 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)
        
        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:]
            
        if self.speaker_embeddings[self.current_speaker]:
            self.mean_embeddings[self.current_speaker] = np.mean(
                self.speaker_embeddings[self.current_speaker], axis=0
            )
        
        return self.current_speaker, similarity
    
    def get_color_for_speaker(self, speaker_id):
        """Return color for speaker ID"""
        if 0 <= speaker_id < len(SPEAKER_COLORS):
            return SPEAKER_COLORS[speaker_id]
        return "#FFFFFF"
    
    def get_status_info(self):
        """Return status information about the speaker change detector"""
        speaker_counts = [len(self.speaker_embeddings[i]) for i in range(self.max_speakers)]
        
        return {
            "current_speaker": self.current_speaker,
            "speaker_counts": speaker_counts,
            "active_speakers": len(self.active_speakers),
            "max_speakers": self.max_speakers,
            "last_similarity": self.last_similarity,
            "threshold": self.change_threshold
        }


class WhisperTranscriber:
    """Simple Whisper transcriber for audio chunks"""
    def __init__(self, model_name="distil-large-v3"):
        self.model = None
        self.processor = None
        self.model_name = model_name
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        
    def load_model(self):
        """Load Whisper model"""
        try:
            from transformers import WhisperProcessor, WhisperForConditionalGeneration
            
            self.processor = WhisperProcessor.from_pretrained(f"distil-whisper/{self.model_name}")
            self.model = WhisperForConditionalGeneration.from_pretrained(f"distil-whisper/{self.model_name}")
            self.model.to(self.device)
            
            return True
        except Exception as e:
            print(f"Error loading Whisper model: {e}")
            return False
    
    def transcribe(self, audio_array, sample_rate=16000):
        """Transcribe audio array"""
        try:
            if self.model is None:
                return ""
            
            # Ensure audio is the right sample rate
            if sample_rate != 16000:
                audio_array = torchaudio.functional.resample(
                    torch.tensor(audio_array).float(),
                    orig_freq=sample_rate,
                    new_freq=16000
                ).numpy()
            
            # Process audio
            inputs = self.processor(audio_array, sampling_rate=16000, return_tensors="pt")
            inputs = inputs.to(self.device)
            
            # Generate transcription
            with torch.no_grad():
                predicted_ids = self.model.generate(inputs["input_features"])
            
            # Decode transcription
            transcription = self.processor.batch_decode(predicted_ids, skip_special_tokens=True)
            
            return transcription[0] if transcription else ""
            
        except Exception as e:
            print(f"Transcription error: {e}")
            return ""


class RealtimeSpeakerDiarization:
    def __init__(self):
        self.encoder = None
        self.audio_processor = None
        self.speaker_detector = None
        self.transcriber = None
        self.audio_buffer = []
        self.processing_thread = None
        self.sentence_queue = queue.Queue()
        self.full_sentences = []
        self.sentence_speakers = []
        self.pending_sentences = []
        self.displayed_text = ""
        self.is_running = False
        self.change_threshold = DEFAULT_CHANGE_THRESHOLD
        self.max_speakers = DEFAULT_MAX_SPEAKERS
        self.audio_chunks = []
        self.chunk_counter = 0
        
    def initialize_models(self):
        """Initialize the speaker encoder and transcription models"""
        try:
            device_str = "cuda" if torch.cuda.is_available() else "cpu"
            print(f"Using device: {device_str}")
            
            # Initialize speaker encoder
            self.encoder = SpeechBrainEncoder(device=device_str)
            encoder_success = self.encoder.load_model()
            
            # Initialize transcriber
            self.transcriber = WhisperTranscriber(FINAL_TRANSCRIPTION_MODEL)
            transcriber_success = self.transcriber.load_model()
            
            if encoder_success and transcriber_success:
                self.audio_processor = AudioProcessor(self.encoder)
                self.speaker_detector = SpeakerChangeDetector(
                    embedding_dim=self.encoder.embedding_dim,
                    change_threshold=self.change_threshold,
                    max_speakers=self.max_speakers
                )
                print("Models loaded successfully!")
                return True
            else:
                print("Failed to load models")
                return False
        except Exception as e:
            print(f"Model initialization error: {e}")
            return False
    
    def process_audio_stream(self, audio_data, sample_rate):
        """Process incoming audio stream data"""
        if not self.is_running or self.encoder is None:
            return
        
        try:
            # Convert audio data to numpy array if needed
            if isinstance(audio_data, tuple):
                sample_rate, audio_array = audio_data
            else:
                audio_array = audio_data
            
            # Ensure audio is float32 and normalized
            if audio_array.dtype != np.float32:
                if audio_array.dtype == np.int16:
                    audio_array = audio_array.astype(np.float32) / 32768.0
                else:
                    audio_array = audio_array.astype(np.float32)
            
            # Ensure mono audio
            if len(audio_array.shape) > 1 and audio_array.shape[1] > 1:
                audio_array = np.mean(audio_array, axis=1)
            
            # Add to buffer
            self.audio_buffer.extend(audio_array.flatten())
            
            # Process when we have enough audio (about 2 seconds)
            target_length = int(sample_rate * 2.0)
            if len(self.audio_buffer) >= target_length:
                self.process_audio_chunk()
                
        except Exception as e:
            print(f"Error processing audio stream: {e}")
    
    def process_audio_chunk(self):
        """Process accumulated audio chunk"""
        try:
            if len(self.audio_buffer) < SAMPLE_RATE:  # Need at least 1 second
                return
            
            # Get audio chunk
            audio_chunk = np.array(self.audio_buffer[:int(SAMPLE_RATE * 2)])
            self.audio_buffer = self.audio_buffer[int(SAMPLE_RATE * 1.5):]  # Keep some overlap
            
            # Transcribe audio
            transcription = self.transcriber.transcribe(audio_chunk, SAMPLE_RATE)
            
            if transcription.strip():
                # Extract speaker embedding
                speaker_embedding = self.audio_processor.extract_embedding(audio_chunk)
                
                # Add to queue for processing
                self.sentence_queue.put((transcription.strip(), speaker_embedding))
                
        except Exception as e:
            print(f"Error processing audio chunk: {e}")
    
    def process_sentence_queue(self):
        """Process sentences in the queue for speaker detection"""
        while self.is_running:
            try:
                text, speaker_embedding = self.sentence_queue.get(timeout=1)
                
                # Store sentence and embedding
                self.full_sentences.append((text, speaker_embedding))
                
                # 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)
                
            except queue.Empty:
                continue
            except Exception as e:
                print(f"Error processing sentence: {e}")
    
    def start_recording(self):
        """Start the recording and transcription process"""
        if self.encoder is None:
            return "Please initialize models first!"
        
        try:
            # Start sentence processing thread
            self.is_running = True
            self.processing_thread = threading.Thread(target=self.process_sentence_queue, daemon=True)
            self.processing_thread.start()
            
            return "Recording started successfully! Start speaking into your microphone."
            
        except Exception as e:
            return f"Error starting recording: {e}"
    
    def stop_recording(self):
        """Stop the recording process"""
        self.is_running = False
        self.audio_buffer = []
        return "Recording stopped!"
    
    def clear_conversation(self):
        """Clear all conversation data"""
        self.full_sentences = []
        self.sentence_speakers = []
        self.pending_sentences = []
        self.displayed_text = ""
        self.audio_buffer = []
        
        if self.speaker_detector:
            self.speaker_detector = SpeakerChangeDetector(
                embedding_dim=self.encoder.embedding_dim,
                change_threshold=self.change_threshold,
                max_speakers=self.max_speakers
            )
        
        return "Conversation cleared!"
    
    def update_settings(self, threshold, max_speakers):
        """Update speaker detection settings"""
        self.change_threshold = threshold
        self.max_speakers = max_speakers
        
        if self.speaker_detector:
            self.speaker_detector.set_change_threshold(threshold)
            self.speaker_detector.set_max_speakers(max_speakers)
        
        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 = "Speaker ?"
                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>')
            
            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}"
    
    def get_status_info(self):
        """Get current status information"""
        if not self.speaker_detector:
            return "Speaker detector not initialized"
        
        try:
            status = self.speaker_detector.get_status_info()
            
            status_lines = [
                f"**Current Speaker:** {status['current_speaker'] + 1}",
                f"**Active Speakers:** {status['active_speakers']} of {status['max_speakers']}",
                f"**Last Similarity:** {status['last_similarity']:.3f}",
                f"**Change Threshold:** {status['threshold']:.2f}",
                f"**Total Sentences:** {len(self.full_sentences)}",
                f"**Audio Buffer Size:** {len(self.audio_buffer)}",
                "",
                "**Speaker Segment Counts:**"
            ]
            
            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]}")
            
            return "\n".join(status_lines)
            
        except Exception as e:
            return f"Error getting status: {e}"


# Global instance
diarization_system = RealtimeSpeakerDiarization()


def initialize_system():
    """Initialize the diarization system"""
    success = diarization_system.initialize_models()
    if success:
        return "✅ System initialized successfully! Models loaded."
    else:
        return "❌ Failed to initialize system. Please check the logs."


def start_recording():
    """Start recording and transcription"""
    return diarization_system.start_recording()


def stop_recording():
    """Stop recording and transcription"""
    return diarization_system.stop_recording()


def clear_conversation():
    """Clear the conversation"""
    return diarization_system.clear_conversation()


def update_settings(threshold, max_speakers):
    """Update system settings"""
    return diarization_system.update_settings(threshold, max_speakers)


def get_conversation():
    """Get the current conversation"""
    return diarization_system.get_formatted_conversation()


def get_status():
    """Get system status"""
    return diarization_system.get_status_info()


def process_audio(audio_data):
    """Process audio from Gradio audio input"""
    if audio_data is not None:
        sample_rate, audio_array = audio_data
        diarization_system.process_audio_stream(audio_array, sample_rate)
    return get_conversation(), get_status()


# Create Gradio interface
def create_interface():
    with gr.Blocks(title="Real-time Speaker Diarization", theme=gr.themes.Monochrome()) as app:
        gr.Markdown("# 🎤 Real-time Speech Recognition with Speaker Diarization")
        gr.Markdown("This app performs real-time speech recognition with automatic speaker identification and color-coding using your browser's microphone.")
        
        with gr.Row():
            with gr.Column(scale=2):
                # Audio input
                audio_input = gr.Audio(
                    source="microphone",
                    type="numpy",
                    streaming=True,
                    label="🎙️ Microphone Input"
                )
                
                # Main conversation display
                conversation_output = gr.HTML(
                    value="<i>Click 'Initialize System' to start...</i>",
                    label="Live Conversation"
                )
                
                # Control buttons
                with gr.Row():
                    init_btn = gr.Button("🔧 Initialize System", variant="secondary")
                    start_btn = gr.Button("🎙️ Start Recording", variant="primary", interactive=False)
                    stop_btn = gr.Button("⏹️ Stop Recording", variant="stop", interactive=False)
                    clear_btn = gr.Button("🗑️ Clear Conversation", interactive=False)
                
                # Status display
                status_output = gr.Textbox(
                    label="System Status",
                    value="System not initialized",
                    lines=10,
                    interactive=False
                )
            
            with gr.Column(scale=1):
                # Settings panel
                gr.Markdown("## ⚙️ Settings")
                
                threshold_slider = gr.Slider(
                    minimum=0.1,
                    maximum=0.95,
                    step=0.05,
                    value=DEFAULT_CHANGE_THRESHOLD,
                    label="Speaker Change Sensitivity",
                    info="Lower values = more sensitive to speaker changes"
                )
                
                max_speakers_slider = gr.Slider(
                    minimum=2,
                    maximum=ABSOLUTE_MAX_SPEAKERS,
                    step=1,
                    value=DEFAULT_MAX_SPEAKERS,
                    label="Maximum Number of Speakers"
                )
                
                update_settings_btn = gr.Button("Update Settings")
                
                # Speaker color legend
                gr.Markdown("## 🎨 Speaker Colors")
                color_info = []
                for i, (color, name) in enumerate(zip(SPEAKER_COLORS, SPEAKER_COLOR_NAMES)):
                    color_info.append(f'<span style="color:{color};">■</span> Speaker {i+1} ({name})')
                
                gr.HTML("<br>".join(color_info[:DEFAULT_MAX_SPEAKERS]))
                
                # Instructions
                gr.Markdown("""
                ## 📋 Instructions
                1. **Initialize System** - Load AI models
                2. **Allow microphone access** when prompted
                3. **Start Recording** - Begin real-time processing
                4. **Speak naturally** - The system will detect different speakers
                5. **Stop Recording** when done
                
                **Note:** Processing happens in real-time with ~2 second chunks for better accuracy.
                """)
        
        # Event handlers
        def on_initialize():
            result = initialize_system()
            if "successfully" in result:
                return (
                    result,
                    gr.update(interactive=True),   # start_btn
                    gr.update(interactive=True),   # clear_btn
                    get_conversation(),
                    get_status()
                )
            else:
                return (
                    result,
                    gr.update(interactive=False),  # start_btn
                    gr.update(interactive=False),  # clear_btn
                    get_conversation(),
                    get_status()
                )
        
        def on_start():
            result = start_recording()
            return (
                result,
                gr.update(interactive=False),  # start_btn
                gr.update(interactive=True),   # stop_btn
            )
        
        def on_stop():
            result = stop_recording()
            return (
                result,
                gr.update(interactive=True),   # start_btn
                gr.update(interactive=False),  # stop_btn
            )
        
        # Connect event handlers
        init_btn.click(
            on_initialize,
            outputs=[status_output, start_btn, clear_btn, conversation_output, status_output]
        )
        
        start_btn.click(
            on_start,
            outputs=[status_output, start_btn, stop_btn]
        )
        
        stop_btn.click(
            on_stop,
            outputs=[status_output, start_btn, stop_btn]
        )
        
        clear_btn.click(
            clear_conversation,
            outputs=[status_output]
        )
        
        update_settings_btn.click(
            update_settings,
            inputs=[threshold_slider, max_speakers_slider],
            outputs=[status_output]
        )
        
        # Process streaming audio
        audio_input.stream(
            process_audio,
            inputs=[audio_input],
            outputs=[conversation_output, status_output],
            time_limit=60,
            stream_every=0.5
        )
        
        # Auto-refresh every 3 seconds
        refresh_timer = gr.Timer(3.0)
        refresh_timer.tick(
            lambda: (get_conversation(), get_status()),
            outputs=[conversation_output, status_output]
        )
    
    return app


if __name__ == "__main__":
    app = create_interface()
    app.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=True
    )