Free Voice Translator

Yes, it is completely free with no signup, no account, and no usage limits. The Whisper, NLLB-200, and Kokoro models run on your own hardware via Hugging Face Transformers.js and kokoro-js instead of a paid cloud API, so there are no per-request costs to pass on. Translate as much speech or text as you want, as often as you want, with no credit card, no API key, and no rate limit.

No. The entire pipeline runs locally in your browser using Transformers.js, kokoro-js, and WebAssembly. After the model files download once, recording, transcription, translation, and speech all happen on your own device with zero network requests carrying your content. Your voice and text are never uploaded, logged, stored, or seen by anyone, which makes this safe for confidential notes and private conversations. You can confirm it by opening the Network tab in your browser DevTools while you translate.

It chains three open-source models: OpenAI Whisper (the onnx-community/whisper-base build) for speech recognition, Meta NLLB-200 distilled 600M (Xenova/nllb-200-distilled-600M) for translation, and Kokoro 82M (onnx-community/Kokoro-82M-v1.0-ONNX) for text to speech. They are loaded through the Transformers.js pipeline() API and kokoro-js, the in-browser equivalents of the Hugging Face Python libraries. Together the quantized weights are several hundred megabytes and download once on first use, then are cached by your browser. They are also shared with the standalone speech-to-text, translator, and text-to-speech tools, so if you have used those, the relevant model is already cached.

You can do either. Record from your microphone or upload an audio file to have Whisper transcribe your speech into the source text, or skip audio entirely and type or paste text directly into the source box. In both cases NLLB-200 then translates that text into your chosen target language. Typing is the fastest path when you already have the text and only need translation and optional speech.

The text-to-speech step uses Kokoro, whose available voices are English-style American and British voices. That means the spoken output sounds most natural for English and other Latin-script target languages and may mispronounce words in other scripts. Speaking the translation is an optional, secondary feature: the reliable core of the tool is the on-screen translated text, which you can always read or copy regardless of the target language.

Translation covers 15 common languages: English, Spanish, French, German, Arabic, Simplified Chinese, Hindi, Portuguese, Russian, Japanese, Korean, Italian, Dutch, Turkish, and Polish, using the FLORES-200 language codes from Meta NLLB-200. Whisper, used for the optional speech step, can recognize speech in many more languages than that, but the translation itself is limited to the language pairs listed in the dropdowns.

The first run includes the one-time download and initialization of the models you use, plus warming up the ONNX Runtime and WebAssembly backend. Transcription and translation also run on your CPU rather than a remote GPU, so they depend on your device. After the weights are cached and the pipelines are warmed up, repeat translations skip the download entirely and start much faster.

Once the models have been downloaded and cached, transcription, translation, and speech run without an internet connection because everything happens on-device. It also works on mobile browsers that support WebAssembly, though phones and tablets are slower than a laptop or desktop, and the first download uses mobile data. For the smoothest experience, load the models once over Wi-Fi and then translate freely afterward.