Real-time voice call pipeline

Orchestrates VAD→STT→LLM→TTS pipeline with barge-in, silence detection, call recording, and ntfy push notifications. The core brain of a real-time voice AI.

"""Per-call pipeline: VAD -> STT -> LLM -> TTS -> audio out."""

import asyncio
import collections
import logging
import os
import re
import threading
import time
import wave
from datetime import datetime
from pathlib import Path
from typing import Callable, Awaitable

import numpy as np

from config import Config
from audio import mulaw_decode, _ULAW_TO_LIN, SileroVAD, SAMPLE_RATE_8K
from stt import SpeechToText
from llm import LLMClient
from tts import TTS
from prompts import get_greeting

log = logging.getLogger(__name__)

# Minimum audio duration (seconds) worth sending to STT.
MIN_AUDIO_DURATION_S = 0.8

# Cooldown after TTS finishes before listening (seconds).
ECHO_COOLDOWN_S = 0.5

# Tag the LLM emits to signal "hang up the call"
HANGUP_TAG = "[HANGUP]"

# --- Barge-in detection con

... (truncated -- full source via MCP)