🎙️ Talkie (ZenithVoice)

An advanced, real-time conversational AI system designed to achieve full-duplex voice interaction. Talkie goes beyond standard Speech-to-Text pipelines by integrating dynamic memory, accent-normalization, and real-time tool execution into a seamless voice-first experience.

🚀 System Capabilities

1. The Core Engines

• ZenithEars (ASR): Powered by Qwen3-ASR with Voice Activity Detection (VAD). Includes a specialized ASR Post-Normalizer to dynamically fix transcription errors for highly specific accents (e.g., Liverpool/Glasgow dialects).
• ZenithBrain (Logic): The orchestration layer handling tool use, goal decomposition, and conversation state.
• ZenithVoiceEngine (TTS): Streams high-fidelity audio chunk-by-chunk using sub-millisecond latent engines (NeuTTS, S1, Kokoro).
• ZenithVerifier (Voice ID): Uses WavLM to cross-reference the speaker's biometric voice against trigger samples in real-time.

2. Deep Conversation Memory (P1.4)

Talkie doesn't just respond; it remembers. It features a fully-tested advanced memory module that can:

• pin_memory: Store specific facts with semantic tags.
• search_memory: Query previous conversation logs.
• compress_context: Token-efficient summarization of long interactions.
• forget_memory: Selective memory wiping on command.

3. Voice-Driven Tool Execution

The system safely maps voice commands to actionable Python tools:

• Retrieving system time
• Web Searching
• Reading specific files
• Executing terminal commands

4. Dynamic Speaker Gate

To prevent the agent from hearing its own voice playing out of the speakers and getting trapped in an infinite feedback loop, Talkie implements a ZenithSpeakerGate. This acts as an intelligent hardware/software barrier, paving the way for upcoming WebRTC Acoustic Echo Cancellation (AEC).

📂 Architecture

• core/ & tools/ - Moduler execution blocks.
• zenith_conductor.py - The main event loop bridging ASR, the LLM Brain, and TTS.
• paul_chat.py & paul_live.py - Specific voice persona implementations featuring regional UK accents ("Paul" / "Ming").
• transformers_compat.py - Custom monkey-patch resolving strict version dependencies in modern transformer libraries.

⚡ Deployment Readiness

Talkie is Production Ready for Phase 1.

# Start the full Voice OS Conductor
python zenith_conductor.py

Note: Phase 2 (True Full-Duplex over audio with multi-threading) is mapped in the system roadmap.