์ค๋ช
Listen.Think.Speak is a fully offline voice chat asset for Windows:
๐๏ธ speech-to-text โ ๐ค LLM โ ๐ฃ๏ธ text-to-speech
Why itโs great
โก Super fast: typical end-to-end replies in 1,500โ4,000 ms (STT โ LLM โ TTS).
๐ 100% local: zero network calls; ideal for offline games and strict privacy.
๐ง Context aware: built-in conversation history for multi-turn dialog.
๐ฃ๏ธ Voices: 28 plug-and-play voice models. [English only for now.]
๐งฉ Drop-in demo: press Load LLM โ Record โ Speak to test in minutes.
Technical details
๐ฅ๏ธ Platform: Windows-only (x64) for now.
๐งฎ Inference: CPU-optimized pipeline for TTS and STT. GPU recommended for LLM.
๐ง Core model: Llama 2 7B (quantized); conversation memory retained per session.
๐ TTS: ONNX voice models with default 22,050 Hz sample rate.
๐๏ธ Model storage: single .bin LLM model file in StreamingAssets; ONNX voices from 60mb.
๐๏ธ Latency breakdown (typical): STT ~300โ800 ms โ LLM ~700โ2,200 ms โ TTS ~300โ1,200 ms โ Total 1.5โ4.0 s. (Tested on intel i7, i9 | Nvidia 2080 Super, 3070-Ti Laptop, 4070-Ti)
Setup in 3 steps
Tools โ Listen.Think.Speak โ Prerequisites โ install components + download the Llama 7B core.
Download one or more ONNX voices, then add them in Speak Manager โ Available Models.
Open the demo, click Load LLM, then Record โ Speak. Done!
Perfect for
๐งโ๐ Immersive NPC conversations
๐ต๏ธ Voice-controlled puzzle/stealth mechanics
๐ฐ Narrative games with dynamic banter
๐ก๏ธ Offline environments
Current support: ๐ช Windows only (macOS/Linux planned).
If you want fast, private, and professional in-game voice interactions without touching the cloud, Listen.Think.Speak is your plug-and-play solution.