Gradium Translates Voice Live — Faster Than GPT

Gradium’s new models translate your speech into French, German, Spanish, or Portuguese in just 3.0 seconds on average, compared to OpenAI’s gpt-realtime-translate which takes 3.6 seconds and Google’s Gemini Live Translate which takes 2.9 seconds. While Gradium falls somewhere in the middle when it comes to latency, it is ahead of the pack in terms of accuracy.

Gradium recently released 2 models, stt-translate for direct text translation and s2s-translate for full-stack speech translation. These 2 models support 20 bi-directional language pairs between English, French, German, Spanish, and Portuguese.

1Why This Approach Is Different

In most cases, voice translation involves multiple steps — speech-to-text (STT) transcription, text-to-text (TT) translation, and then text-to-speech (TTS) playback in the target language. Each step introduces latency and error rates, and any missteps during STT can throw off subsequent steps.

Gradium simplifies this process with 2 models instead of 3. With stt-translate, you get STT and TT translation in one step without the middleman. It’s also all done over a single connection, which keeps latency low while maintaining accuracy. And with s2s-translate, you have the option to choose a target voice or copy an existing voice, which isn’t possible with OpenAI’s realtime translation.

According to Gradium, their models outperform gpt-realtime-translate and gemini-3.5-live-translate in terms of BLEU score (a common metric used to evaluate the quality of a machine translation), and also outperform GPT in MetricX. Of course, these are self-reported results, so take them with a grain of salt, but at least the architecture is sound.

2What This Means If You're Building

Whether you’re building a voice assistant for meetings, a language learning app, a multi-lingual customer support bot, or experimenting with live translation, now you have another choice besides OpenAI and Google. Gradium’s API is available at gradium.ai/translate. You can use their Python SDK with async streaming, which means you don’t have to change anything about your current workflow.