Skip to main content

Translation Engine Settings

This page contains four configuration cards in order: Text Translation Engine Settings, Voice Translation Engine Settings, ASR Settings (speech recognition), and TTS Settings (speech synthesis). The entry points for setting the source and target languages of each engine are as follows:

  • OCR (text) translation source/target language: Configured in the "Translation Language Settings" of OCR Image Translation Settings.
  • Voice translation recognition language and target language: Configured in the "Speaker / Microphone Voice Translation Settings" of Voice Translation Settings.

The source language can be set to auto-detect, so you don't need to manually adjust it when switching game languages; the target language is usually set to your native language (such as Simplified Chinese).

Text Translation Engine Settings​

Configure the engine used for OCR screenshot text translation.

Translation Mode​

Select the processing channel for translation requests:

  • Built-in Translation Service: Directly use the service integration built into the project, ready to use out of the box without applying for development keys on the platform.
  • Custom Configuration: Suitable for users who have already applied for relevant platform API keys on their own, or have deployed inference models locally.
  • Local Translation: Use the bundled Hy-MT2 local translation model for fully offline translation, with no internet connection or API keys required. The local model loads on demand and automatically unloads after 30 seconds of inactivity to free memory.
    • After selecting "Local", online settings such as vendor / authentication are hidden, leaving only the Inference Device and Inference Core Count (see "Inference Device and Core Count" below).

Service Type​

Select which type of translation engine to use (shown when not in "Local" mode):

  • General Translation Service: Traditional "word-to-word / sentence-to-sentence" machine translation with fast response speed. Includes mainstream services such as Tencent, Microsoft, Volcano Engine, Alibaba, Baidu, Youdao, Google, DeepL, and Huawei. (Youdao Translation and Huawei Translation are shown depending on availability.)
  • LLM Large Model Service: Based on large language models, better at context understanding and tone polishing. Includes services such as OpenAI, DeepSeek, Alibaba Bailian (Tongyi Qianwen), Volcano Engine (Doubao), SiliconFlow, Anthropic, Google Gemini, Tencent Hunyuan, Kimi, MiniMax, Baidu Qianfan, xAI, and Zhipu.

Preferred Vendor and Model​

  • Preferred Vendor: Select from the vendor list corresponding to the service type above.
  • Preferred Model: Only valid for LLM services. In built-in mode, the available model list for that vendor is automatically fetched for selection; in custom mode, you can manually enter the model name.

Service Credentials​

  • Built-in Service Authentication: In built-in mode, some vendors require you to fill in the auth username and auth password assigned by the service provider (leave blank to use the built-in default key).
  • Custom Service Authentication: In custom mode, according to the selected service type, the following input items are displayed (in the following order):
    • API ID: Some service providers (such as Baidu Translate) require providing an application ID at the same time, fill it in here.
    • API KEY: Fill in the API key generated by the corresponding service provider.
    • Model Name: Only valid for LLM mode. Fill in the specific model you want to use, such as deepseek-chat, gpt-4o-mini, glm-4.
    • API URL: Only valid for LLM mode. If you use locally deployed Ollama, LM Studio, or proxy relay services, you can fill in the custom interface address here (such as http://localhost:11434/v1/chat/completions); leave blank to use the default interface of the selected service provider.
    • Prompt: Only valid for LLM mode. Used to customize the translation style and requirements of the model. The default prompt is "You are a translation expert, the text to be translated is the result of OCR recognition...".

Voice Translation Engine Settings​

Configure the translation engine used for voice translation.

Translation Mode​

  • Local (default): Use the bundled Hy-MT2 local translation model for offline translation.
  • Custom: Use your own API keys to access third-party translation services.

Note: Voice translation does not support the built-in translation service mode; only local and custom modes are supported. Voice translation uses an independent vendor / credential configuration.

Service Type and Vendor​

Same as text translation, supporting General Translation Service and LLM Large Model Service, with the same vendor list. Voice translation's custom credentials (API ID / API KEY / Model Name / API URL / Prompt) are saved independently from text translation.

Voice Translation Default Prompt​

The default LLM prompt for voice translation is: "You are a voice translation expert, the input is continuous speech-recognized segments...", which differs from the default prompt for text translation.

Inference Device and Core Count​

After selecting the "Local" mode, the following two items are displayed (consistent with the settings of the same name in the ASR / TTS cards):

  • Inference Device: CPU (default) or a graphics card enumerated by the subprocess via Vulkan (integrated / discrete, showing the actual device name).
  • Inference Core Count: The number of CPU inference cores used (range 1 to the device core count, default 4). This slider is automatically hidden after selecting a GPU.

ASR Settings​

Configure the speech recognition (automatic speech recognition) engine. This card was moved from the "Voice Translation Settings" page and contains all options that are strongly tied to the model.

Recognition Mode​

  • Local (default): Use the bundled local Qwen3-ASR model for offline recognition, without loading online services.
  • Custom: Connect to an online recognition service (currently Alibaba Bailian DashScope is supported), without loading the local model. After selecting, the following are displayed:
    • Vendor: Currently Alibaba Bailian DashScope (the dropdown reserves room for expansion).
    • Model Name: Divided into "real-time" and "non-real-time" presets according to the endpoint protocol, so you can directly select a preset model; the last item "Custom" lets you manually enter the model name and API URL.
    • API URL: In custom mode, you can enter a custom interface address (leave blank to use the default address).
    • API Key: The key for the online service. When left blank, the interface shows an "API Key not filled in" prompt.

Local-Mode-Only Options​

  • Speech Recognition Model: Select the tier of the local recognition model.
    • Fast (0.6B) (default): Faster response, uses less memory.
    • High-Precision (1.7B): Higher accuracy, uses more memory (requires about 3 GB of available memory). Switching to high-precision first pops up a confirmation prompt, and takes effect only after confirmation.
  • Inference Device: CPU (default) or an enumerated GPU device.
  • Inference Core Count: The number of CPU inference cores (1 to the device core count, default 4); hidden after selecting a GPU.
  • Simulated Streaming Output: When enabled, intermediate recognition results are output incrementally at roughly 1-second intervals, making the display closer to real-time subtitles; when disabled, the output is produced in one go only after the entire utterance is recognized. Enabling it first pops up a confirmation prompt.

TTS Settings​

Configure the speech synthesis (text-to-speech) engine, used to read aloud the translation of the microphone channel. This card was moved from the "Voice Translation Settings" page.

Synthesis Mode​

  • Local (default): Use the bundled local Qwen3-TTS model for offline synthesis.
  • Custom: Connect to an online synthesis service (currently Alibaba Bailian DashScope is supported). After selecting, the Vendor, Model Name (including "real-time / non-real-time" presets and a "Custom" manual entry), API URL, and API Key are displayed.

Local-Mode-Only Options​

  • TTS Real-time Streaming Playback: When enabled, playback happens while inference is running, resulting in lower latency; when disabled, playback occurs all at once after complete synthesis.
  • Inference Device: CPU (default) or an enumerated GPU device. A confirmation prompt pops up before switching to GPU (GPU inference currently has an issue of abnormally increasing VRAM usage, possibly occupying about 6.5 GB of VRAM).
  • Voice: Default (Random), Female A, Female B, Male A, Male B, Custom (Clone). Switching the voice only updates the engine configuration without reloading the model, and takes effect on the next read-aloud.
  • Voice Cloning: Click "Microphone Clone" or "Speaker Clone" to capture a human voice; the interface shows the number of seconds captured, and after about 10 seconds in total it is automatically saved as a "Custom (Clone)" voice; during capture you can click "Stop Capture" to interrupt. Please use a clean, continuous single-person voice.
  • Inference Core Count: The number of CPU inference cores (1 to the device core count, default 4); hidden after selecting a GPU.

If synthesis fails, the error message returned by the online synthesis will be displayed in red text within the card, making it easier to troubleshoot API Key, quota, or network issues.

Service Status Prompt​

When selecting "Built-in" mode, if the system is loading the available interface list, a rotating loading animation will be displayed next to the dropdown box. If the request fails due to network issues, it will be displayed in red countdown text, at which point you can wait for a retry or switch to "Custom" mode.

In built-in mode, some services that are temporarily unavailable or lack credentials will be marked as disabled / prompted in the interface.