Duplex: Multi-Language Duplex Chat via LLM

📢 News

• 2026-07-13 🖼️ Unified UI style system – StyleManager now manages fonts, colors, and widget styles across the main UI, control panel, and dialogs, with 14 preset themes.
• 2026-07-13 🔧 Control panel refactor – Improved QTabBar style alignment, WA_TranslucentBackground transparency handling, and Fusion style integration to ensure stylesheet consistency.
• 2026-07-11 📦 Duplex package refactor – Centralized configuration with Pydantic, migrated settings.json to a nested structure, and completed migration from Windows code to Linux/Arch.
• 2026-07-11 📥 Automatic changelog system – Dynamically fetches release history from GitHub Releases, removing the local CHANGELOG dependency.
• 2026-07-10 🔊 Groq Whisper ASR – Added Groq Whisper API using whisper-large-v3, eliminating the need for local ASR models.
• 2026-07-09 🌐 Remote Whisper engine – Added WebSocket bridge to remote Whisper servers with support for private deployment.
• 2026-07-08 🔧 ASR subprocess architecture – Moved ASR model loading into an independent subprocess to reduce main process memory usage; workers automatically restart after crashes.
• 2026-07-08 📦 Model proxy support – Unified proxy handling for Silero, ModelScope, and HuggingFace model downloads.
• 2026-07-07 🎯 SenseVoice optimization – Added fp16 CUDA acceleration and optimized VAD inference with inference_mode.
• 2026-07-06 💾 Local faster-whisper support – Added local GGUF / CTranslate2 model loading with automatic cache layout detection.
• 2026-07-05 🧹 Translation context optimization – Fixed context management and improved model cache decision logic.
• 2026-07-04 🎨 Transcription persistence – Added translation history persistence with memory protection and MEM diagnostic logging.
• 2026-07-03 📝 Text export feature – Added export support for original text, translated text, or all transcripts to .txt files via tray menu and overlay menu.
• 2026-07-02 🎭 Anime-Whisper integration – Integrated litagin/anime-whisper for Japanese anime / galgame speech recognition.
• 2026-07-01 🔀 Dual-pipeline threading – Separated audio capture and ASR processing threads to prevent long sentence recognition from blocking RMS/VAD updates.
• 2026-07-01 ✨ Translation prompt optimization – Added ASR error correction, fluency rules, preset selectors, and proper noun preservation.
• 2026-06-30 🔄 Incremental ASR pipeline – Added multi-language sentence splitting with pysbd; incremental output reduces long audio latency.
• 2026-06-29 🏠 Subtitle window mode – Added standalone transparent subtitle window for OBS capture, using QPainterPath outlined text, auto wrapping, and entrance/exit animations.
• 2026-06-28 🎛️ Control panel system – Added 7 configuration tabs: VAD/ASR, translation, style, subtitle, benchmark, cache, and changelog.
• 2026-06-27 🎨 Theme system – Added 14 UI presets including Dracula, Nord, Monokai, Solarized, Gruvbox, Tokyo Night, Catppuccin, One Dark, Everforest, and Kanagawa.

Earlier news

• 2026-06-25 🏷️ v0.4.0 release — Rebranded to duplex, introduced streaming UI and high-contrast default style.
• 2026-06-22 📦 uv packaging workflow — Added GitHub Actions release pipeline and automated package builds.
• 2026-06-20 🐛 Installation hardening — Restricted supported versions to Python 3.10-3.12, added SSL fallback for Silero downloads.
• 2026-06-18 🔧 FunASR Nano integration — Preserved punctuation and added pre-downloaded Qwen3-0.6B weights.
• 2026-06-15 📊 Changelog system update — Added May features and historical records.
• 2026-06-10 🎙️ SenseVoice integration — Added FunASR SenseVoice backend with GPU-accelerated Japanese recognition.
• 2026-06-05 🧠 Translator refactor — Built unified OpenAI-compatible API client with streaming responses and JSON Schema.
• 2026-05-28 ⚡ Memory optimization — Added ASR input padding buckets and cache reuse to reduce memory pressure.
• 2026-05-20 🔊 FunASR integration — Added local ASR model support with multi-engine switching.
• 2026-05-15 🎯 VAD overhaul — Added Silero VAD + energy detection dual mode, backtrack split, and adaptive silence thresholds.
• 2026-05-10 🎨 Style system — Added 14 preset themes and independent original/translation font configuration.
• 2026-05-01 🚀 Initial version — Started LiveTrans project with PyAudio capture, faster-whisper, and OpenAI translation.

📦 Install

sudo pacman -S python-pip python-venv pyqt6 portaudio
git clone https://github.com/your-repo/duplex.git

🚀 Quick Start

cd duplex && uv sync && source .venv/bin/activate && pip install -e . && duplex

// control panel model config in settings.json
"models": [
  {
    "name": "Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled",
    "api_base": "http://127.0.0.1:8001/v1",
    "api_key": "try",
    "model": "Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled",
    "proxy": "none",
    "streaming": true,
    "no_system_role": false,
    "no_think": true,
    "json_response": false,
    "context_turns": 0,
    "input_price": 0.0,
    "output_price": 0.0,
    "overrides": {},
    "extra_body": {}
  }
],

✨ Features

📝 Subtitle Modes

Mode	Description	Best For
🎮 Overlay Mode	Transparent floating subtitle window with high contrast style. Supports dragging, click-through, and always-on-top behavior.	Gaming, livestreaming, real-time communication
🎬 Subtitle Window Mode	Independent transparent subtitle window optimized for OBS capture, with animated outlined text effects.	Recording, streaming, video production

---

🎙️ ASR Engines

Engine	Description	Features
⚡ faster-whisper	CTranslate2 optimized Whisper implementation	Local inference, supports GGUF / CTranslate2 models
🇨🇳 FunASR SenseVoice	Alibaba SenseVoice speech recognition model	GPU accelerated, excellent Japanese recognition
🪶 FunASR Nano	Lightweight models based on Qwen3-0.6B and similar architectures	Local inference, CPU-friendly
🎌 Anime-Whisper	Specialized Whisper model by litagin/anime-whisper	Optimized for anime / galgame audio
☁️ Groq Whisper	Cloud-based whisper-large-v3 inference	No local model required, ultra-low latency
🌐 Remote Whisper	WebSocket-based remote ASR bridge	Supports private/self-hosted deployment

---

🌍 Translation System

• 🔌 OpenAI-compatible API

  • OpenAI
  • DeepSeek
  • Qwen
  • Groq
  • Other compatible providers

• ⚡ Real-time Streaming Translation

  • Incremental response streaming
  • Live translation progress display

• 🧩 Structured Output

  • JSON Schema constrained responses
  • Reliable machine-readable output

• 🧠 Context Awareness

  • Multi-turn conversation context
  • Model-specific parameter overrides:
    • temperature
    • top_p
    • max_tokens
    • etc.

• 💰 Usage Tracking

  • Token counting
  • API cost estimation
  • Per-model pricing configuration

---

🎤 Voice Activity Detection (VAD)

Feature	Description
🔥 Silero VAD	ONNX-based real-time voice activity detection with WASM/CUDA acceleration
🔋 Energy Detection	Lightweight fallback detector for low-resource environments
📊 Adaptive Silence	Dynamic silence threshold using P75 × 1.2 (0.3s ~ 2.0s range)
✂️ Backtrack Split	Finds the lowest energy valley when splitting long speech segments
📈 Progressive Silence	Adaptive pause handling: • <3s → full buffer • 3-6s → half buffer • 6-10s → quarter buffer