By Nicole Nguyen

I don't speak Spanish, but there I was, chattering away like I had spoken it my entire life.

I was testing a new real-time translator coming to Google Meet video chats. The results, while not perfect, show how good AI is getting at processing language and mimicking human emotion.

Google announced the speech translator at the company's annual I/O keynote Tuesday, and I got to experience it for myself a few days earlier, live on a video call with two Spanish speakers.

It isn't just that the translator turns your words into another language. It emulates your voice and tone, translating with a few seconds of lag. The effect is like watching an overdubbed foreign-language speaker on a news broadcast -- but the voice-over is created by AI in the speaker's same voice.

A Google Meet pop-up warned me and the two Spanish-speaking employees, Cami and Jair, that the experimental translation might not always be correct. We clicked to agree, then began to converse with each other in our native languages.

They talked about where they like to eat after work and travel for weekend getaways in various Latin American locales. Their digitally produced English alter egos had slight Spanish accents. For the most part, the translation was fluid, with minimal lag.

The feature can work with up to 100 participants, though even with just us three, there was some confusing crosstalk due to the delay. As a speaker, you don't hear your translated voice -- so you don't know when it stops talking.

There were times when the audio was initially stilted, like there was a connection issue, but eventually the translation caught up.

Deciding how much of a speaker's audio to translate at a time was one of the team's biggest hurdles, says Awaneesh Verma, Google's senior director of real-time communication.

The technology starts interpreting before it has the full message. That's hard work, because of context. If you say "bear," you could mean an animal, giving birth or carrying something.

When I tested it with my husband Will, who spoke Spanish, it translated the English word "match," as in a tennis match, to "fight" in Spanish. He also said whenever I started speaking, the first sentence was a bit garbled, but it smoothed out after that.

Sometimes the voice-over placed an emphasis in the wrong place or produced broken English: "The heat...the climate...always very warm." And some direct translations just don't sound right: "I am fascinated by the power to have many options." Translation is an art and Google's beta isn't flawless, but I got the gist.

I realized how good this clone tech was once I heard my own voice, while watching a playback. It sounded scarily like me. Even Will was impressed.

I was describing my next trip to Spain for a big family party thrown by my sister-in-law. My colleague, who speaks Spanish, listened to it after and said the audio was a "mixed bag," especially at first. The accent, she said, wasn't native. (Funny: It also gave Will a slight Spanish accent when translating his voice into English.)

Speech translation will start rolling out to Google's advanced Gemini subscribers ($20 a month and up) starting today. English and Spanish are available first, with Italian, German and Portuguese in the works. Options for business customers will come later this year. Google says it doesn't store any data from the meeting when translation is used, and it doesn't use your voice or any other meeting audio to train its AI models.

While there's much room for improvement, the potential is exciting for anyone who has ever had to speak in a language that isn't their own.

I know first hand: I recently lived in Paris and often felt I couldn't be nearly as funny, interesting or expressive in French. This could break down some of those barriers, as long as nothing gets lost in translation.

Write to Nicole Nguyen at nicole.nguyen@wsj.com

(END) Dow Jones Newswires

May 20, 2025 13:29 ET (17:29 GMT)

Copyright (c) 2025 Dow Jones & Company, Inc.