How to use the Conversation Translator
A free tool for short everyday sentences, with clear limits for data, devices and language.
Which languages work?
German ↔ English and English ↔ Hindi use direct models. German ↔ Hindi uses English as an intermediate step. This pivot translation is experimental and may lose greetings, nuance or information. Hindi is a specific language; other Indian languages are not included here.
What happens on first use?
After confirmation, your browser downloads public model files from Hugging Face, about 110 MB per direct direction or 220 MB for German ↔ Hindi. The interface also loads about 0.9 MB of JavaScript and 21.6 MB of WASM for the first translation. No paid inference API is called. Your browser may cache files; deletion, storage limits and offline availability depend on it.
Does it work on my smartphone?
The interface adapts to smartphones and translation uses WebAssembly rather than a desktop-only browser API. Six language directions were actually run in Chrome on a Windows desktop. Android, iOS, Safari, memory use and speed on real mobile devices have not yet been tested. If a model cannot load, try a short direct German/English translation in a recent browser on Wi-Fi.
How do recording and reading aloud work?
Record uses the speech recognition offered by your browser. You first accept the possible audio data flow to the browser provider and grant microphone access. The transcript remains editable. Missing recognition, permission or a service does not prevent typing. Reading aloud uses only matching voices that your browser reports as local device voices. When none is available, reading aloud stays disabled.
What does the page store?
Inputs, translations and conversation history stay in this tab’s memory. When you deliberately change the interface language, the draft is transferred once through the same tab’s session storage and removed after restoration. Download as text lets you save a local text file. Clear removes the current input and reply; Delete history also removes the conversation list. Model files may remain in the browser cache independently.
What are the limits?
Use short sentences with one clear thought, and check numbers, names and negations. Translations may omit details or misunderstand them. Medical, legal, official or safety-critical conversations require a suitable language professional. Quality, universal support and unrestricted offline use are not guaranteed.
Costs and support
The conversation tool requires no account, subscription or paid language model. Your existing device, electricity and data costs are separate. Static hosting avoids additional technical fixed operating costs; normal hosting limits apply. A later labelled sponsor or separately commissioned setup service is possible. Practical usability is not proof of demand or income.
Sources and attribution
Local text translation: Transformers.js 3.8.1 and quantised OPUS-MT models (Marian, WASM). ONNX conversions: Xenova. Original models: Helsinki-NLP. German → English, English → Hindi and Hindi → English: Apache 2.0. English → German: CC BY 4.0. Models can make mistakes; this interface does not replace review by people who know the languages.
Transformers.js · DE → EN · Apache 2.0 · EN → DE · CC BY 4.0 · EN → HI · Apache 2.0 · HI → EN · Apache 2.0 · Xenova ONNX · Browser speech recognition
Optional advertising-partner offer. All free content and downloads remain available independently.
Open the advertising-partner offer ↗Optional Adsterra advertisement. Advertising loads only after your agreement; it may process your IP address, device information and identifiers and use cookies. All content and downloads stay freely available without agreement.