Arif: Urdu Voice Assistant (JARVIS-style)

Current focus: This repo has ONE goal — build "Arif", a voice assistant that takes spoken Urdu commands and performs actions on the user's Windows PC (file/folder operations, closing apps, etc.), then replies with Urdu voice. Only work on Arif.

⬇️ ⬇️ ⬇️ CLICK ME ⬇️ ⬇️ ⬇️

➡️ ➡️ ➡️           ⬅️ ⬅️ ⬅️

⬆️ ⬆️ ⬆️                       ⬆️ ⬆️ ⬆️

About the user / interaction style

• The user speaks Urdu (romanized meaning like "falani folder mein chale jao", "falana folder bana do", "meri saari files close kar do").
• The wake word / assistant name is "Arif".
• Arif replies back in Urdu voice.
• Platform: Windows 11, primary language Python. Shell is PowerShell.

Pipeline (voice command → action → voice reply)

[Mic] → [Wake word "Arif"] → [Record until silence (VAD)]
     → [STT + translate Urdu→English]   (Whisper)
     → [Intent parsing → structured action JSON]   (Groq LLM)
     → [Confirmation gate for risky actions]
     → [Executor runs the action on Windows]
     → [Compose Urdu reply] → [TTS Piper → .wav] → [Play audio]
     → loop

Stage details

1. Wake word — detect "Arif" before recording (openWakeWord / Porcupine), so the system isn't parsing everything it hears.
2. Record — capture mic audio; use Voice Activity Detection (VAD, e.g. webrtcvad or silero-vad) to stop recording when the user stops speaking.
3. STT + translation — convert Urdu speech directly to English text.
   • Whisper handles this in one step via its translate task (non-English speech → English text). Options: local faster-whisper (GPU) OR Groq's hosted whisper-large-v3 audio endpoint (much faster, no local compute). Prefer Groq's audio API to keep latency low.
4. Intent parsing — send the English text to Groq (chat completions). Groq must return a structured JSON action (use JSON mode / tool-calling), NOT free prose, so the executor can act reliably. See "Action schema" below.
5. Confirmation gate — destructive/ambiguous actions (delete, overwrite, move, close apps, actions spanning 2+ files) require a spoken Urdu confirmation before running. Read-only/create actions can run directly.
6. Executor — perform the action with Python: os, shutil, pathlib, subprocess, psutil (list/kill processes), pygetwindow/pywin32 (windows).
7. TTS — synthesize the Urdu reply with Piper (ur_PK-fasih-medium) to a .wav, then play it. Working code already validated (see below).

Action schema (Groq output contract)

Groq should return an object the executor can dispatch on, e.g.:

{
  "action": "create_folder | delete | move | copy | open_app | close_app | close_all | list | open_folder | none",
  "args": { "path": "...", "target": "...", "name": "..." },
  "needs_confirmation": true,
  "reply_urdu": "<short Urdu sentence to speak back>"
}

Keep the list of valid action values and their args as the single source of truth the executor and the Groq system prompt both reference.

Safety rules (important)

• Confirm before destructive actions: delete, overwrite, move, mass-close apps, or anything touching 2+ files.
• Allowlist directories: restrict file operations to user-approved roots; refuse paths outside them.
• Never run shell/exec content that came from the transcription without validation.
• Log every executed action (timestamp, transcript, action, result) for auditability.
• Support a dry-run mode during development.

Known-good building blocks

Piper TTS (Urdu) — verified working

from piper import PiperVoice
import wave

model_path = "ur_PK-fasih-medium.onnx"
config_path = "ur_PK-fasih-medium.onnx.json"
voice = PiperVoice.load(model_path, config_path=config_path)

text = "آپ کا کیا حال ہے؟"
with wave.open("saul_urdu_response.wav", "wb") as wav_file:
    voice.synthesize_wav(text, wav_file)

Tech stack (planned)

• Python (Windows)
• Mic capture: sounddevice / pyaudio; VAD: webrtcvad or silero-vad
• Wake word: openWakeWord or pvporcupine
• STT+translate: Groq whisper-large-v3 audio API (or local faster-whisper)
• LLM intent: Groq chat completions (JSON mode / tool calling)
• TTS: Piper (ur_PK-fasih-medium)
• Audio playback: sounddevice / simpleaudio
• System actions: os, shutil, pathlib, subprocess, psutil, pygetwindow

Config / secrets

• Groq API key via environment variable (e.g. GROQ_API_KEY) or .env — never hardcode.

Conventions

• Keep the pipeline modular: separate files/modules per stage (wake, stt, intent, executor, tts) so each can be tested in isolation.
• Test each stage standalone before wiring the full loop.

Status

• ✓ Done Piper Urdu TTS validated
• ✓ Done Project scaffolded: config.py, recorder.py, stt.py, intent.py, executor.py, tts.py, main.py
• ✓ Done Isolated venv at .venv/ created with --system-site-packages (reuses the machine's existing torch==1.12.1+cu113 install instead of downloading a second copy; do NOT pip install into the global Python directly -- it's shared with other tools like aider-chat and got its versions clobbered once already, then restored)
• ✓ Done STT moved off Groq entirely -- runs locally via HF openai/whisper-small on the GTX 1050 (task="translate", Urdu audio -> English text in one step, confirmed on cuda:0). Only intent.py still calls Groq (for the LLM intent-parsing step), to conserve API usage. transformers/tokenizers are pinned venv-local (4.40.0 / 0.19.1) because the global transformers install is independently broken (tokenizers/protobuf/tensorflow version rot from other tools) -- stt.py sets USE_TF=0, USE_FLAX=0, and PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python to route around that global rot.
• ✓ Done Groq JSON-mode intent parsing and executor smoke-tested individually and pass

Run it

D:\JARVIS-ARIF\.venv\Scripts\python.exe main.py

Hold F9, speak in Urdu, release. Ctrl+C to quit (You can say Allah Hafiz/Khuda Hafiz too😁).