Google has expanded its Gemini model family with the introduction of Gemini 3.8 Flash TTS and Flash-Lite TTS. These text-to-speech models are designed to handle more than 100 languages. According to a report from The Decoder, the primary feature of the Flash TTS model is the ability to generate entirely new synthetic voices based on descriptive text prompts. Additionally, the models allow for more granular control over audio output, such as including stage directions for specific lines of dialogue and producing conversations between two distinct voices from a single input script.
Advancements in Audio Personalization
The release of these models suggests a shift toward more customizable and expressive AI-generated audio. While traditional text-to-speech systems often rely on a fixed set of pre-recorded voice profiles, the Flash TTS model’s ability to interpret text descriptions for voice design allows for a broader range of characterization. The inclusion of a voice cloning feature—which requires a 30-second audio sample to build a profile—places Google’s new tools alongside existing cloning technologies in the industry. By supporting over 100 languages at launch, Google is positioning these models for global applications, potentially ranging from localized content creation to more interactive virtual assistants that can switch between different personas or handle complex scripts involving multiple speakers.
Implementation and Safety Considerations
The announcement did not specify the exact release date for these models or whether they will be integrated directly into the Gemini chatbot interface or offered primarily through API access for developers. It is also unclear what specific safety measures or watermarking technologies Google will implement to prevent the misuse of the 30-second voice cloning feature. Furthermore, the report did not provide details regarding the computational requirements for running these models or if there will be a tiered pricing structure for the Flash-Lite version compared to the standard Flash TTS model. As these tools become available, the industry will likely monitor how well the stage directions feature handles complex emotional cues or non-verbal sounds.
These new TTS models represent a technical step in Google’s efforts to make synthetic speech more adaptable and easy to produce from simple text inputs.
Lead image: Andreu Marquès / Pexels
