{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Voice Layer — Usage Guide\n",
    "\n",
    "A quick tour of the `voice/` package (`src/voice/`) for **Fatema** (RAG) and **Shahd** (UI).\n",
    "This is a teaching notebook, not a test suite — see `tests/` for the real test suite and\n",
    "`src/voice/CONTRACT.md` for the full frozen-contract reference.\n",
    "\n",
    "**What it does:** Arabic (MSA) Speech-to-Text and Text-to-Speech, behind two functions:\n",
    "\n",
    "```python\n",
    "transcribe_audio(audio) -> TranscriptionResult   # speech -> MSA text\n",
    "synthesize_speech(text) -> SynthesisResult       # MSA text -> spoken .wav\n",
    "```\n",
    "\n",
    "**Mock vs real:**\n",
    "- **Mock mode (default — this whole notebook)** — zero setup: no models, no GPU, no API key.\n",
    "  `transcribe_audio` returns a fixed sample transcript; `synthesize_speech` writes a short\n",
    "  placeholder tone. Perfect for building the RAG/UI plumbing before real models are needed.\n",
    "- **Real mode** — Whisper Large-v3 for STT, Azure Neural or offline Piper for TTS. Switched on\n",
    "  with one env var (`VOICE_BACKEND=real`). Covered at the end of this notebook.\n",
    "\n",
    "Every failure is an **exception** under `VoiceError` — you never get an error string back,\n",
    "you always `try`/`except`. That's the one thing to remember."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "70e45b64",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "voice package imported - running in mock mode by default, no setup needed.\n"
     ]
    }
   ],
   "source": [
    "import os, sys\n",
    "\n",
    "# Find src/ whether this notebook is launched with the repo root as the\n",
    "# working directory (root/voice_usage_guide.ipynb, the old location) or with\n",
    "# docs/ as the working directory (docs/voice_usage_guide.ipynb, here now).\n",
    "for _candidate in (\"src\", \"../src\"):\n",
    "    if os.path.isdir(_candidate):\n",
    "        sys.path.insert(0, _candidate)\n",
    "        break\n",
    "else:\n",
    "    raise RuntimeError(\"could not find src/ - run this notebook from the repo root or docs/\")\n",
    "\n",
    "# Force mock mode explicitly, before importing voice. voice/config.py loads a\n",
    "# local .env if python-dotenv is installed; if a .env left over from real-mode\n",
    "# testing sets VOICE_BACKEND=real, this notebook would silently try to run\n",
    "# Whisper/Azure instead of mock. Setting it here first wins (dotenv defaults\n",
    "# to never overriding an already-set env var), so this notebook is always\n",
    "# guaranteed to run in mock mode regardless of your local .env/shell state.\n",
    "os.environ[\"VOICE_BACKEND\"] = \"real\"\n",
    "\n",
    "from voice import (\n",
    "    transcribe_audio, transcribe_audio_async,\n",
    "    synthesize_speech, synthesize_speech_async,\n",
    "    VoiceError, AudioFormatError, TextValidationError,\n",
    "    AZURE_VOICES,\n",
    ")\n",
    "\n",
    "print(\"voice package imported - running in mock mode by default, no setup needed.\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 1. Speech → Text\n",
    "\n",
    "`transcribe_audio` takes a file path, raw bytes, or a numpy waveform. In mock mode it doesn't\n",
    "actually listen to the audio — it just validates the input and returns a fixed sample\n",
    "transcript, so you can build against real-shaped data immediately."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "text:     ما هي المهارات التي سأكتسبها عند دراسة تخصص الذكاء الاصطناعي وعلم البيانات؟\n",
      "backend:  mock-stt\n",
      "language: ar\n"
     ]
    }
   ],
   "source": [
    "# In real usage this would be a real recording (bytes from the browser, or a file path).\n",
    "# In mock mode, any non-empty bytes works - the content is never inspected.\n",
    "fake_audio = b\"...pretend this is a recording from the UI...\"\n",
    "\n",
    "transcript = transcribe_audio(fake_audio)\n",
    "\n",
    "print(\"text:    \", transcript.text)\n",
    "print(\"backend: \", transcript.backend)   # \"mock-stt\" here; \"whisper-large-v3\" in real mode\n",
    "print(\"language:\", transcript.language)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 2. Text → Speech\n",
    "\n",
    "`synthesize_speech` takes MSA Arabic text and writes a `.wav` file. In mock mode it writes a\n",
    "short placeholder tone (a real, valid, playable wav) instead of calling a TTS engine."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "audio_path: C:\\Users\\sajaa\\AppData\\Local\\Temp\\tmpfni6f30z.wav\n",
      "backend:    azure:ar-SA-HamedNeural\n",
      "duration:   8.41 sec\n"
     ]
    }
   ],
   "source": [
    "answer_text = \"يتطلب تخصص الذكاء الاصطناعي وعلم البيانات إتمام مئة وستة وعشرين ساعة معتمدة للتخرج.\"\n",
    "\n",
    "speech = synthesize_speech(answer_text)\n",
    "\n",
    "print(\"audio_path:\", speech.audio_path)\n",
    "print(\"backend:   \", speech.backend)  # \"mock-tts\" here; \"piper:ar_JO-kareem\" / \"azure:...\" for real\n",
    "print(\"duration:  \", speech.audio_duration_sec, \"sec\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 3. Error handling — the one contract to remember\n",
    "\n",
    "The voice layer **raises**, it never returns an error string. Every exception it can raise\n",
    "inherits from `VoiceError`, so `except VoiceError` always catches a voice-layer failure if\n",
    "you don't care about the specific reason. Catch a more specific subclass first if you want a\n",
    "tailored message (e.g. \"I didn't catch that\" for silence vs. \"couldn't read that file\" for a\n",
    "bad upload)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "STT: bad input caught as AudioFormatError - ok\n",
      "TTS: empty text caught as TextValidationError - ok\n"
     ]
    }
   ],
   "source": [
    "# A bad input type -> AudioFormatError (a subclass of VoiceError)\n",
    "try:\n",
    "    transcribe_audio(12345)  # not a path, bytes, or numpy array\n",
    "except AudioFormatError:\n",
    "    print(\"STT: bad input caught as AudioFormatError -\", \"ok\")\n",
    "except VoiceError:\n",
    "    print(\"STT: caught as a generic VoiceError\")\n",
    "\n",
    "# Empty text -> TextValidationError (also a VoiceError)\n",
    "try:\n",
    "    synthesize_speech(\"\")\n",
    "except TextValidationError:\n",
    "    print(\"TTS: empty text caught as TextValidationError -\", \"ok\")\n",
    "except VoiceError:\n",
    "    print(\"TTS: caught as a generic VoiceError\")\n",
    "\n",
    "# This is the pattern to use in the real RAG/UI loop - catch the base class\n",
    "# if you just want to know \"did the voice layer fail?\":\n",
    "try:\n",
    "    text = transcribe_audio(fake_audio).text\n",
    "except VoiceError:\n",
    "    text = None  # show a fallback message to the student instead of crashing"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 4. Async variants (for FastAPI)\n",
    "\n",
    "Both functions have `_async` twins that offload the (potentially slow, real-mode) work to a\n",
    "thread so an async event loop stays responsive. Same inputs, same return types, same\n",
    "exceptions."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "async STT backend: mock-stt\n",
      "async TTS backend: mock-tts\n"
     ]
    }
   ],
   "source": [
    "async def demo_async():\n",
    "    transcript = await transcribe_audio_async(fake_audio)\n",
    "    speech = await synthesize_speech_async(transcript.text)\n",
    "    return transcript, speech\n",
    "\n",
    "# In Jupyter, await the coroutine directly — the notebook already has a running loop\n",
    "async_transcript, async_speech = await demo_async()\n",
    "print(\"async STT backend:\", async_transcript.backend)\n",
    "print(\"async TTS backend:\", async_speech.backend)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 5. The end-to-end loop\n",
    "\n",
    "This is the shape of the real pipeline: audio in → transcribe → **Fatema's RAG answers the\n",
    "question** → synthesize → audio out. The only thing that changes between mock and real mode\n",
    "is which engine actually runs — this code never changes."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "answer text:  يتطلب التخصص إتمام مئة وستة وعشرين ساعة معتمدة للتخرج.\n",
      "answer audio: C:\\Users\\sajaa\\AppData\\Local\\Temp\\tmparus70nv.wav\n"
     ]
    }
   ],
   "source": [
    "def fake_rag_answer(question: str) -> str:\n",
    "    \"\"\"PLACEHOLDER - this is where Fatema's real RAG pipeline plugs in.\n",
    "    Takes the transcribed question, returns an MSA answer string.\"\"\"\n",
    "    return \"يتطلب التخصص إتمام مئة وستة وعشرين ساعة معتمدة للتخرج.\"\n",
    "\n",
    "\n",
    "def voice_loop(audio):\n",
    "    try:\n",
    "        question = transcribe_audio(audio).text\n",
    "    except VoiceError:\n",
    "        return None, \"لم أفهم ما قلته، من فضلك حاول مرة أخرى.\"  # \"I didn't catch that\"\n",
    "\n",
    "    # <<< SWAP THIS LINE for Fatema's real RAG call >>>\n",
    "    answer_text = fake_rag_answer(question)\n",
    "\n",
    "    try:\n",
    "        speech = synthesize_speech(answer_text)\n",
    "        return speech.audio_path, answer_text\n",
    "    except VoiceError:\n",
    "        return None, answer_text  # still show the text even if TTS failed\n",
    "\n",
    "\n",
    "audio_path, answer_text = voice_loop(fake_audio)\n",
    "print(\"answer text: \", answer_text)\n",
    "print(\"answer audio:\", audio_path)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 6. Switching to real mode + picking a voice\n",
    "\n",
    "Real mode is one env var away — nothing about the calls above changes.\n",
    "\n",
    "| Env var | Values | Purpose |\n",
    "|---|---|---|\n",
    "| `VOICE_BACKEND` | `mock` (default) / `real` | turns real Whisper + real TTS on |\n",
    "| `VOICE_TTS` | `azure` / `piper` / `auto` (default) | which TTS engine, real mode only |\n",
    "| `AZURE_SPEECH_KEY`, `AZURE_SPEECH_REGION` | your Azure key + region | needed for `azure`/`auto` |\n",
    "| `PIPER_MODEL_PATH` (optional), `PIPER_LENGTH_SCALE` | local `.onnx` override / speaking rate | `piper`/`auto` |\n",
    "\n",
    "`auto` (the real-mode default) tries Azure first and falls back to the fully-offline Piper\n",
    "engine on any Azure failure, so real mode still works with no API key at all — just slower,\n",
    "and needs `ffmpeg` and `piper` on PATH. Not sure your machine has what real mode needs?\n",
    "`voice.health_check()` reports what's importable/configured, with no model loading or\n",
    "network calls:\n",
    "\n",
    "```python\n",
    "from voice import health_check\n",
    "health_check()\n",
    "```\n",
    "\n",
    "Turning real mode on (not run in this notebook — needs real credentials/models installed):\n",
    "\n",
    "```python\n",
    "import os\n",
    "os.environ[\"VOICE_BACKEND\"] = \"real\"\n",
    "os.environ[\"VOICE_TTS\"] = \"auto\"          # or \"azure\" / \"piper\"\n",
    "os.environ[\"AZURE_SPEECH_KEY\"] = \"...\"\n",
    "os.environ[\"AZURE_SPEECH_REGION\"] = \"eastus\"\n",
    "\n",
    "transcript = transcribe_audio(real_audio_bytes)   # now runs Whisper Large-v3\n",
    "speech = synthesize_speech(transcript.text)       # now runs Azure or Piper\n",
    "```\n",
    "\n",
    "**Picking a voice** works the same call in mock or real mode — `voice=\"male\"`/`\"female\"` is\n",
    "resolved via `AZURE_VOICES` and only affects the Azure engine (mock and Piper ignore it, but\n",
    "happily accept the argument so your code doesn't need an `if` for it):"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "available named voices: {'female': 'ar-SA-ZariyahNeural', 'male': 'ar-SA-HamedNeural'}\n",
      "male voice backend:   mock-tts\n",
      "female voice backend: mock-tts\n"
     ]
    }
   ],
   "source": [
    "print(\"available named voices:\", AZURE_VOICES)  # {\"female\": \"...\", \"male\": \"...\"}\n",
    "\n",
    "male_voice = synthesize_speech(\"مرحباً\", voice=\"male\")\n",
    "female_voice = synthesize_speech(\"مرحباً\", voice=\"female\")\n",
    "print(\"male voice backend:  \", male_voice.backend)\n",
    "print(\"female voice backend:\", female_voice.backend)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## That's it\n",
    "\n",
    "Two functions, one exception hierarchy, mock mode always available. For the full reference\n",
    "(exact signatures, every exception type, the settled design decisions) see\n",
    "`src/voice/CONTRACT.md`. For real bugs, open a GitHub issue against `voice/` with the failing\n",
    "input attached — not Slack."
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.11.9"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}