CLAUDE.md

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

LiveTrans is a real-time audio translation system for video players on Windows. It captures system audio via WASAPI loopback, runs speech recognition, and translates via LLM APIs, displaying results in a transparent overlay.

Current phase: Phase 0 Python prototype (Phase 1 will be a C++ DirectShow Audio Tap Filter).

Running

# Must use the project venv (system Python lacks dependencies)
.venv/Scripts/python.exe main.py

Linter: ruff (installed globally). Run python -m ruff check --select F,E,W --ignore E501,E402 *.py to lint. E402 is intentionally ignored because main.py requires torch before PyQt6.

Architecture

The pipeline runs in a background thread: Audio Capture (32ms chunks) -> VAD -> ASR -> Translation (async) -> Overlay

main.py (LiveTransApp)
  |-- model_manager.py     Centralized model detection, download, cache utils
  |-- audio_capture.py     WASAPI loopback via pyaudiowpatch, auto-reconnects on device change
  |-- vad_processor.py     Silero VAD / energy-based / disabled modes, progressive silence + backtrack split
  |-- asr_engine.py        faster-whisper (Whisper) backend
  |-- asr_sensevoice.py    FunASR SenseVoice backend (better for Japanese)
  |-- asr_funasr_nano.py   FunASR Nano backend
  |-- translator.py        OpenAI-compatible API client, streaming, make_openai_client()
  |-- subtitle_overlay.py  PyQt6 transparent overlay (2-row header: controls + model/lang combos)
  |-- control_panel.py     Settings UI (5 tabs: VAD/ASR, Translation, Style, Benchmark, Cache)
  |-- dialogs.py           Setup wizard, model download/load dialogs, ModelEditDialog
  |-- benchmark.py         Translation benchmark (BENCH_SENTENCES, run_benchmark())
  |-- log_window.py        Real-time log viewer

Threading Model

• Main thread: Qt event loop (all UI)
• Pipeline thread: _pipeline_loop in LiveTransApp - reads audio, runs VAD/ASR/translation
• ASR loading: Background thread via _switch_asr_engine (heavy model load, ~3-8s)
• Cross-thread UI updates use Qt signals (e.g., add_message_signal, update_translation_signal)
• ASR readiness tracked by _asr_ready flag; pipeline drops segments while loading

Configuration

• config.yaml - Base configuration (audio, ASR, translation, subtitle defaults)
• user_settings.json - Runtime settings persisted by control panel (models, VAD params, ASR engine choice, optional cache_path). Takes priority over config.yaml on load.

Model Config

Each model in user_settings.json has: name, api_base, api_key, model, proxy ("none"/"system"/custom URL), optional no_system_role (bool, for APIs that reject system messages like Qwen-MT).

Proxy handling: proxy="none" uses httpx.Client(trust_env=False) to bypass system proxy; proxy="system" uses default httpx behavior (env vars).

Overlay UI (subtitle_overlay.py)

DragHandle is a 2-row header bar:

• Row 1: Draggable title + action buttons (Paused/Running, Clear, Settings, Monitor, Quit)
• Row 2: Checkboxes (Click-through, Top-most, Auto-scroll) + Model combo + Target Language combo

Style system:

• DEFAULT_STYLE and STYLE_PRESETS defined in subtitle_overlay.py — 14 presets including terminal themes (Dracula, Nord, Monokai, Solarized, Gruvbox, Tokyo Night, Catppuccin, One Dark, Everforest, Kanagawa)
• Default style is high-contrast (pure black background, white translation text, 14pt)
• Original and translation text have independent font_family fields (original_font_family, translation_font_family)
• SubtitleOverlay.apply_style(style) updates container/header backgrounds, window opacity, and rebuilds all message HTML
• Style dict stored in user_settings.json under "style" key; forwarded via settings_changed signal → main.py → overlay.apply_style()
• Backward compat: old font_family key auto-migrated to split fields in apply_style()

Key overlay features:

• Top-most: Toggles WindowStaysOnTopHint; requires setWindowFlags() + show() to take effect
• Click-through: Uses Win32 WS_EX_TRANSPARENT on the scroll area while keeping header interactive
• Auto-scroll: Controls whether new messages/translations auto-scroll to bottom
• Model combo: Populated from user_settings.json models list; switching emits model_switch_requested signal
• Target Language combo: Emits target_language_changed; synced from settings on startup

Settings UX

• Auto-save with debounce: All control panel settings (combos, spinboxes) auto-save after 300ms debounce via _auto_save() → _do_auto_save(). No manual Save button needed.
• Slider special handling: VAD/Energy sliders update labels in real-time but only trigger save on sliderReleased (mouse) or immediately for keyboard input (isSliderDown() check).
• Apply Prompt button: Kept because TextEdit shouldn't trigger on every keystroke. Also persists to disk.
• Cache path: Default ./models/ (not ~/.cache). Applied at startup in main.py before import torch via model_manager.apply_cache_env().

Startup Flow

1. main.py reads user_settings.json and calls apply_cache_env() before import torch
2. First launch (no user_settings.json) → SetupWizardDialog: choose hub + path + download Silero+SenseVoice
3. Non-first launch but models missing → ModelDownloadDialog: auto-download missing models
4. All models ready → create main UI (overlay, panel, pipeline)
5. Runtime ASR engine switch: if uncached → ModelDownloadDialog, then _ModelLoadDialog for GPU loading

Key Patterns

• torch must be imported before PyQt6 to avoid DLL conflicts on Windows (PyTorch 2.9.0+ bug, see main.py and pytorch#166628)
• Cache env vars set at module level in main.py before import torch to ensure TORCH_HOME is respected
• Deferred initialization: ASR model loading and settings application happen via QTimer.singleShot(100) after UI is shown to prevent startup freeze
• make_openai_client() in translator.py is the single shared function for proxy-aware OpenAI client creation (used by both translator and benchmark)
• create_app_icon() in main.py generates the app icon; set globally via app.setWindowIcon() so all windows inherit it
• Model cache detection (is_asr_cached, get_local_model_path) checks both ModelScope and HuggingFace paths to avoid redundant downloads when switching hubs
• Settings log output (_apply_settings) filters out models and system_prompt to avoid leaking API keys
• FunASR Nano: asr_funasr_nano.py does os.chdir(model_dir) before AutoModel() so relative paths in config.yaml (e.g. Qwen3-0.6B) resolve locally instead of triggering HuggingFace Hub network requests
• Translator defaults to 10s timeout via make_openai_client() to prevent API calls from hanging indefinitely
• Log window is created at startup but hidden; shown via tray menu "Show Log"
• Audio chunk duration is 32ms (512 samples at 16kHz), matching Silero VAD's native window size for minimal latency
• VAD adaptive silence mode: tracks recent pause durations, sets silence threshold to P75 × 1.2, auto-adjusts between 0.3s~2.0s
• VAD progressive silence: buffer越长接受越短的停顿切分 (<3s=full, 3-6s=half, 6-10s=quarter of silence_limit)
• VAD backtrack split: max duration时回溯confidence history找最低谷切分，remainder保留到下一段；三级策略（绝对低谷→相对低谷20%→低于均值兜底）
• FunASR disable_pbar=True required in all generate() calls — tqdm crashes in GUI process when flushing stderr
• ASR engine lifecycle: each engine exposes unload() (move to CPU + release) and to_device(device) (in-place migration). Device switching uses to_device() for PyTorch engines (SenseVoice/FunASR) and full reload for ctranslate2 (Whisper). Release order: unload() → del → gc.collect() → torch.cuda.empty_cache()
• Whisper (ctranslate2) only accepts device="cuda" not "cuda:0"; device index passed via device_index param. Parsed from combo text like "cuda:0 (RTX 4090)" in _switch_asr_engine
• VAD speech density filter: _flush_segment() discards segments where <25% of chunks are above confidence threshold (noise rejection)
• ASR text density filter: segments ≥2s producing ≤3 alnum characters are discarded as noise
• Settings file uses atomic write (write to .tmp then os.replace) to prevent corruption on crash
• stop() joins pipeline thread before flushing VAD to prevent concurrent _process_segment calls
• Cancelled ASR download/failed load restores _asr_ready if old engine is still available
• Translator._build_system_prompt catches format errors in user prompt templates, falls back to DEFAULT_PROMPT

Language & Style

• Respond in Chinese
• Code comments in English only where critical
• Commit messages without Co-Authored-By