Google's new AI models can create voices from text prompts
What's the story
Google has launched two advanced text-to-speech (TTS) models, Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, on its cloud platform. The models come with similar application programming interfaces (APIs), making it easy for developers to use them together. While Flash-Lite TTS focuses on cost-efficiency and speed, the more expensive Flash TTS delivers superior audio quality, ideal for tasks like audiobook creation.
Model capabilities
Custom voices and parameters
The Gemini models come with a library of over 2,000 prepackaged voices. Developers can also create custom voices using natural language prompts.
Google allows customization of parameters such as an AI speaker's vocal timbre, accent, and pacing.
Another way to customize these models is by generating a synthetic voice from a 30-second audio sample (with the speaker's consent).
Security features
Audio watermarks for transparency
Google uses a technology called SynthID to embed an audio watermark in AI-generated speech. The watermark is inaudible to humans but detectable by AI detection tools.
Additionally, the company attaches a C2PA record to every generated audio file, detailing when it was created, if it has been modified since, and related information. This ensures transparency and security for users of the new models.
Performance evaluation
Competing models on multiple language-specific versions of Voice Arena
The new Gemini models were evaluated using an audio quality benchmark developed by Hume AI Inc.
Flash TTS and Flash-Lite TTS topped the chart, outperforming several competing models on multiple language-specific versions of Voice Arena. This is a testament to their superior output quality based on human feedback.
Google hopes these advancements will enhance user experiences in products like Gemini Notebook and Google Vids.