Engines & Models: Choosing and Configuring Local vs. Cloud
TransBox only provides the client and adapter layer. It does not resell or host any third-party large-model service.
For speech recognition (ASR), machine translation (MT), and speech synthesis (TTS), you can choose Local Offline or Cloud Service under Settings, then finish the download or enter an API key as prompted.
1. Factory Defaults (matching the current client)
| Stage | Default engine | Notes |
|---|---|---|
| Speech recognition | Local Qwen3-ASR | Fully local inference; data never leaves your machine |
| Machine translation | Local Tencent Hunyuan Hy-MT2-1.8B | Officially preferred local MT model for interpretation (about 3.6 GB) |
| Speech synthesis | Local Kokoro-82M | Lightweight local TTS with low first-packet latency |
If local weights are not downloaded yet, opening Translation / Read aloud will prompt you to download first. You can also switch to cloud in one click in Settings and enter the vendor API key.
2. Local Offline Models
2.1 What They Offer
- After a one-time download, recognition / translation / read-aloud still work with the network cable unplugged (depending on the selected engine)
- Suitable for privacy-sensitive scenarios, unstable networks, and long-term high-frequency use
- In testing, local Tencent Hunyuan Hy-MT2-1.8B delivered the best latency and translation quality; local models are recommended first
2.2 Download and Storage
- Open Settings → Speech Recognition / Machine Translation / Speech Synthesis → Local Offline Model
- Review the hardware rating and disk usage, then click Download Model
- You can change the path / open the folder / clear cache
Default cache locations (Windows example):
- ModelScope:
%USERPROFILE%\.cache\modelscope\models\... - Hugging Face:
%USERPROFILE%\.cache\huggingface\hub\...
2.3 Download Mirrors
You can switch mirrors on the settings card:
| Mirror | Best for |
|---|---|
| Global official nodes (Hugging Face) | Prefer when overseas or when the network is reliable |
| High-speed domestic mirror (ModelScope / domestic CDN) | More stable on networks in mainland China |
If the global node fails, the client automatically falls back to the domestic backup mirror. You can also select the domestic source first by hand.
2.4 Local Proxy (Optional)
Affects model downloads only, and is off by default:
- Check Enable proxy download
- The default address hint is
http://127.0.0.1:7897(port varies by tools such as Clash — set it for your machine) - Uncheck to disable
- Domestic users on the “High-speed domestic mirror” usually do not need a proxy
- Overseas users generally should not enable it
- TransBox will not force the local proxy when the option is unchecked
2.5 Hardware and Performance
- A dedicated NVIDIA GPU is recommended for local ASR / MT. Without one, fall back to CPU (slower) or use cloud
- For local MT (Hy-MT2-1.8B), about 3.6 GB+ VRAM is recommended for smoother performance
- Actual frame-level latency and accuracy depend on acoustics, speaking rate, and glossaries — no performance guarantee
3. Cloud Services (Bring Your Own API Key)
3.1 How It Works
| Point | Detail |
|---|---|
| Who provides the capability | Third-party cloud vendors (e.g. Alibaba Cloud Bailian, Tencent Hunyuan, Xiaomi MiMo, and others) |
| Who pays | You enable the service, top up, and obtain an API key in the vendor console |
| TransBox’s role | Protocol adaptation, settings UI, and calls only; does not collect or resell vendor quota |
| Key storage | Saved in local configuration; never written into Workshop packages or uploaded to TransBox servers |
3.2 Configuration Steps (General)
- Register on the vendor site, enable the model service, and create an API key
- In Settings, choose Cloud Service and the specific engine
- Paste the key; adjust Base URL / model name if needed
- Click Test Connection
- Click Save & Apply
3.3 Rate Limits and Concurrency
Cloud calls are subject to the vendor account’s RPM / TPM quotas (usually aggregated by master account — not “only one connection allowed”).
Default personal-tier quotas are typically enough for a single user; sharing one key across people shares the quota. On 429 / rate-limit errors:
- Retry later, or reduce simultaneous Translation + Read aloud load
- Check quota in the vendor console, request a raise, or upgrade to a business tier
- This is an upstream limit, not a TransBox client fault
3.4 Privacy
- Local engines: audio and text are processed on your machine (depending on plugin implementation)
- Cloud engines: related content is sent to the third-party API you configured — read that vendor’s privacy policy
- Do not enable cloud recognition/translation for highly confidential meetings until you understand the data flow
4. Compatibility and Responsibility Boundaries (Read Carefully)
- TransBox is client software: Steam purchases grant use of this software. Local open-source weights and cloud APIs are neither produced nor hosted by TransBox.
- Models and vendors may change: names, pricing, rate limits, regions, and auth methods are decided by third parties. We sync what we can in release notes, but cannot guarantee 100% long-term compatibility with every upstream.
- Scope of adapters: officially built-in engines are whatever appears in Settings. Third-party MOD engines are maintained by community authors — use at your own risk.
- No service-availability guarantee: local model downloads depend on the network and mirrors; cloud depends on vendor SLAs. Interpretation latency and accuracy depend on acoustics, languages, glossaries, hardware, and more — no result is promised.
- Costs: local weight downloads may incur data charges; cloud usage is billed by the vendor and is unrelated to the one-time software price.
- Lawful use: comply with local laws and each cloud vendor’s terms. Do not use this software for illegal interception, infringing redistribution, or similar purposes.
- Support scope: we help with client logs, default engines, and the integrations listed in the official docs. For vendor account setup, billing, bans, or leaked keys, contact that vendor first.
5. How Should I Choose?
| Your situation | Recommendation |
|---|---|
| Privacy matters / no stable internet | Choose local everywhere; download weights first |
| Need the highest recognition accuracy and can go online | Use cloud for ASR; local or cloud for MT/TTS |
| VRAM < 4 GB | Prefer cloud, or load fewer local models at once |
| Slow downloads in mainland China | Pick the high-speed domestic mirror; if still slow, try the local proxy |
| Shared key for a team / multiple instances | Raise the vendor quota, or use separate keys per person |