backend/voice/audio_ws.py

# backend/voice/audio_ws.py
import asyncio
import json
import logging
import numpy as np
import websockets
import wave
import io
import os
from datetime import datetime

from .vad import SileroVAD
from .stt import STTPipeline
from .tts import TTSPipeline
from .wake_word import WakeWordDetector
from backend.ws.agent_ws import ws_manager

import sys
parent_dir = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
if parent_dir not in sys.path:
    sys.path.insert(0, parent_dir)
from modules.assistant_identity import set_mode
# Shared pipeline lock — also used by space_injection.py so wake-word utterances
# and space-ack speech cannot physically overlap.
from backend.voice.space_injection import get_pipeline_lock

AUDIO_LOGS_DIR = os.path.join(os.path.dirname(__file__), "audio_logs")
os.makedirs(AUDIO_LOGS_DIR, exist_ok=True)
TRANSCRIPT_LOG_PATH = os.path.join(AUDIO_LOGS_DIR, "transcripts_log.txt")

active_audio_ws = None

class AudioWebSocketServer:
    def __init__(self, host="127.0.0.1", port=8767):
        global active_audio_ws
        active_audio_ws = self
        self.host = host
        self.port = port
        self.vad = SileroVAD()
        self.stt = STTPipeline()
        self.tts = TTSPipeline()
        self.wake_word = WakeWordDetector()
        self.stop_future = None
        self.interrupt_flag = False
        # Use the shared pipeline lock so space_injection cannot fire mid-utterance
        self.wake_lock = get_pipeline_lock()

    def interrupt(self):
        self.interrupt_flag = True

    async def handler(self, websocket):
        logging.info("Client connected to Voice WebSocket.")
        
        speech_buffer = bytearray()
        is_speaking = False
        silence_chunks = 0
        MAX_SILENCE_CHUNKS = 15 # Approx 0.5s of silence to trigger end of utterance
        
        is_tts_playing = False
        
        try:
            async for message in websocket:
                if self.interrupt_flag:
                    speech_buffer.clear()
                    is_speaking = False
                    silence_chunks = 0
                    self.interrupt_flag = False
                    
                if isinstance(message, bytes):
                    if is_tts_playing:
                        continue # Mute mic input while TTS is speaking to prevent self-triggering
                    audio_np = np.frombuffer(message, dtype=np.int16).astype(np.float32) / 32768.0
                    
                    # 1. Wake Word Detection (openWakeWord)
                    keyword = self.wake_word.process_frame(audio_np)
                    if keyword:
                        async with self.wake_lock:
                            if "friday" in keyword:
                                set_mode("friday")
                            elif "jarvis" in keyword:
                                set_mode("jarvis")
                            logging.info(f"openWakeWord triggered: {keyword}")
                            await ws_manager.broadcast({"event": "voice:wake_word", "payload": {"agent": keyword}})
                            speech_buffer = bytearray() # Clear buffer to start listening
                            is_speaking = True
                            silence_chunks = 0
                        continue
                    
                    # 2. VAD Confirmation
                    vad_res = self.vad.process_chunk(audio_np)
                    if vad_res:
                        is_speaking = True
                        silence_chunks = 0
                        speech_buffer.extend(message)
                    elif is_speaking:
                        speech_buffer.extend(message)
                        silence_chunks += 1
                        
                        if silence_chunks >= MAX_SILENCE_CHUNKS:
                            is_speaking = False
                            silence_chunks = 0
                            
                            if len(speech_buffer) > 16000:
                                logging.info("Utterance complete. Transcribing...")
                                wav_io = io.BytesIO()
                                with wave.open(wav_io, 'wb') as wf:
                                    wf.setnchannels(1)
                                    wf.setsampwidth(2)
                                    wf.setframerate(16000)
                                    wf.writeframes(speech_buffer)
                                    
                                # 3. Whisper Transcription
                                text, detected_lang = await self.stt.transcribe(wav_io.getvalue())
                                if text.strip():
                                    timestamp = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
                                    wav_filepath = os.path.join(AUDIO_LOGS_DIR, f"{timestamp}.wav")
                                    with open(wav_filepath, "wb") as f:
                                        f.write(wav_io.getvalue())
                                        
                                    await websocket.send(json.dumps({
                                        "type": "transcript",
                                        "text": text.strip(),
                                        "language": detected_lang
                                    }))
                                    
                                    # --- JARVIS 10X Gaming Coach Interrupt ---
                                    # If the user yells a tactical interrupt, kill the audio immediately.
                                    interrupt_phrases = ["quiet", "stop", "shut up", "enemy on me", "contact"]
                                    if any(phrase in text.lower() for phrase in interrupt_phrases):
                                        logging.warning(f"JARVIS 10X Tactical Interrupt Detected: '{text.strip()}'")
                                        # Send hard kill signal to frontend audio player
                                        await websocket.send(json.dumps({"type": "audio_interrupt"}))
                                        # Skip routing to LLM for this utterance unless it's a query
                                        if "what went wrong" not in text.lower():
                                            speech_buffer = bytearray()
                                            continue
                                    
                                    # 4. Route to LLM
                                    logging.info(f"Routing to ReAct LLM Agent: {text.strip()} (Lang: {detected_lang})")
                                    await ws_manager.handle_client_event(
                                        websocket=None,
                                        data={"type": "chat", "text": text.strip(), "language": detected_lang}
                                    )
                            speech_buffer = bytearray()

                elif isinstance(message, str):
                    try:
                        data = json.loads(message)
                        msg_type = data.get("type")
                        
                        if msg_type == "tts_request":
                            text = data.get("text", "")
                            personality = data.get("personality", "jarvis")
                            language = data.get("language", "en")
                            
                            if text:
                                logging.info(f"Synthesizing TTS for: {text} in {language}")
                                wav_bytes = await self.tts.synthesize(text, personality, language)
                                
                                async def _play_and_mute():
                                    nonlocal is_tts_playing
                                    is_tts_playing = True
                                    await websocket.send(json.dumps({
                                        "type": "tts_response_start",
                                        "text": text
                                    }))
                                    await websocket.send(wav_bytes)
                                    await websocket.send(json.dumps({
                                        "type": "tts_response_end"
                                    }))
                                    # 16-bit PCM = 2 bytes per sample @ 24kHz
                                    audio_duration = len(wav_bytes) / 24000.0 / 2.0
                                    await asyncio.sleep(audio_duration + 0.5) # +0.5s trailing silence buffer
                                    is_tts_playing = False
                                
                                asyncio.create_task(_play_and_mute())
                    except json.JSONDecodeError:
                        pass

        except websockets.exceptions.ConnectionClosed:
            logging.info("Client disconnected from Voice WebSocket.")
        except Exception as e:
            logging.error(f"Voice WebSocket Error: {e}")

    async def start_server(self):
        self.stop_future = asyncio.Future()
        server = None
        for p in range(self.port, self.port + 11):
            try:
                server = await websockets.serve(self.handler, self.host, p)
                self.port = p
                logging.info(f"Voice WebSocket running on ws://{self.host}:{self.port}")
                break
            except OSError as e:
                if "address already in use" in str(e).lower() or "10048" in str(e):
                    logging.warning(f"Voice WS Port {p} in use, trying next...")
                    continue
                raise
        
        if server:
            await self.stop_future  # run until stopped
            server.close()
            await server.wait_closed()
            logging.info("Voice WebSocket server shut down cleanly.")

    def stop(self):
        if self.stop_future and not self.stop_future.done():
            self.stop_future.set_result(True)

def run_in_thread(host="127.0.0.1", port=8767):
    import os
    if os.environ.get("CLOUD_ENV", "false").lower() == "true":
        logging.info("[AudioWS] Cloud mode — skipping local audio WebSocket server (no mic on HF Spaces).")
        return None
    server = AudioWebSocketServer(host=host, port=port)
    def _run():
        loop = asyncio.new_event_loop()
        asyncio.set_event_loop(loop)
        loop.run_until_complete(server.start_server())
        
    import threading
    t = threading.Thread(target=_run, daemon=True)
    t.start()
    return server

if __name__ == "__main__":
    logging.basicConfig(level=logging.INFO)
    asyncio.run(AudioWebSocketServer().start_server())