Article contents
- Introduction to the new real-time translator
- Breaking language barriers in real time
- Integration across various platforms and devices
- Diffusion Gemma text generation model
- Balancing performance speed with quality
- Frequently asked questions
Introduction to the new real-time translator
In what is considered one of the company’s boldest and most ambitious technical initiatives in deploying locally integrated artificial intelligence technologies on smart devices, Google officially announced on Tuesday the launch of its advanced real-time voice translation tool named “Gemini 3.5 Live Translate”. This revolutionary new model relies on generative artificial intelligence capabilities to process and translate spoken conversations directly and almost instantly, supporting a broad list of more than 70 different global languages. This launch underscores the company’s leadership in the fierce race to develop artificial intelligence models capable of transforming daily human communication and making it more seamless. On the same day filled with major technical announcements, the company’s advanced research unit, known as DeepMind, unveiled an open-source experimental model called “Diffusion Gemma”, specifically designed to generate programming and written text at speeds up to four times faster than traditional auto-regressive models known on the market.
Breaking language barriers in real time
The new voice model departs radically and innovatively from the operating mechanisms of older traditional translation systems that rely on a turn-taking principle, where those slow systems required the speaker to stop talking and wait for seconds until the system finished processing the sentence and producing the corresponding translation. Instead of this disjointed approach, the new system processes the flow of spoken speech and translates it continuously and automatically without interruption, lagging behind the actual speaker by only a few short seconds to ensure contextual accuracy. The technical achievement here is not limited to speed alone; the system also features an accurate and impressive ability to preserve the speaker’s natural vocal intonation, respect the original delivery speed, and maintain fundamental pitch levels. These combined features make the final voice output sound natural and familiar to the human ear, significantly enhancing the quality and depth of direct human communication between individuals from different cultures and linguistic backgrounds.
Integration across various platforms and devices
To ensure this advanced technology reaches the widest possible segment of users, the intelligent model is being rolled out gradually across several key company interfaces. First, the model has become available worldwide through the new update of the official translation app dedicated to smartphones running Android and iOS operating systems. Second, the system is available in a public preview version aimed at developers and programmers through a dedicated application programming interface and via the AI Studio platform. Third, starting this current month, the company has begun providing a special preview version to a select number of enterprise customers subscribed to workspace services by integrating the technology directly into the virtual meetings platform. This deep integration into the meetings platform expands the scope of voice translation capabilities immensely and remarkably, rising from just 5 previously supported languages to over 70 currently supported languages, enabling the creation of more than 2,000 different reciprocal language combinations within a single interactive meeting. The company also added a new and convenient feature to the mobile operating system known as listening mode, which allows users to simply listen to direct and continuous translations through the phone’s standard earpiece without needing additional external headphones. To ensure the highest degrees of transparency and reliability, all artificially generated audio clips are marked with a subtle, invisible digital watermark known as “SynthID” to help verify content produced by AI tools.
Diffusion Gemma text generation model
Along related lines and within the same technical momentum, the company’s research unit released a large language model based on the mixture-of-experts architecture containing approximately 26 billion parameters, named “Diffusion Gemma”. This open-source experimental model generates text using an engineering approach very similar to how artificial intelligence image models create and compose visual canvases. The system starts from a state of random digital noise and very quickly and in parallel refines and clarifies entire blocks of text up to 256 tokens at once, rather than the old, time-consuming mechanism of sequentially predicting one word at a time mathematically. This massive technical system was built and designed on the strong and solid architecture of the Gemma 4 model, and despite its large structural size, it cleverly relies on activating only 3.8 billion parameters during the actual inference and generation process, significantly and tangibly improving computational resource consumption efficiency.
Balancing performance speed with quality
Thanks to these precise architectural improvements, this experimental model achieves exceptional and massive processing speeds exceeding 1,000 tokens per second when running on an advanced enterprise-grade processing card, and reaches about 700 tokens per second when using consumer-grade graphics cards aimed at individual users. The weights of this advanced model have been made available to the general public under an open-source license, encouraging the broader tech community to explore its capabilities and develop new applications. The company’s CEO, Mr. Sundar Altman, praised this important and distinctive technical achievement, describing the new system as being like a thoroughbred racehorse achieving inference speeds four times faster. With all these impressive achievements and developments, the leading tech company issued an explicit technical warning to developers that this model is still in its early, incomplete experimental stage and technically lags behind the standard and more stable model regarding absolute text output quality standards. The company recommended using this model primarily and exclusively for rapid local tasks where responsive speed is a critically decisive factor that tolerates no delay, such as direct and quick code editing operations, rather than complex final production deployments requiring the highest levels of linguistic accuracy.
Frequently asked questions
Question: What is the most prominent feature of the newly launched real-time translator?
Answer: The new translator features the ability to provide continuous and direct real-time voice translation for over 70 different languages without needing to wait for the speaker to stop talking.
Question: How does the innovative fast text generation technology work?
Answer: The technology relies on a mechanism of generating and clarifying entire grouped blocks of text in parallel at the same time using an engineering process called diffusion, which boosts processing speed by fourfold.
Question: Can users listen to direct translation in a practical and convenient manner?
Answer: Yes, through the new listening mode which allows holding the mobile phone close to the ear and hearing the translation clearly, using it just like regular voice phone calls without needing extra headphones.
Question: Is the new text generation model considered ready for final and reliable commercial use?
Answer: The model is still in its early incomplete experimental stage, and the company strongly advises using it for programming and local tasks that require extremely high speed while avoiding it in sensitive projects requiring absolute accuracy.
