Google Expands Audio Capabilities with New Gemini TTS Models
Google has launched two dedicated Gemini 3.8 Flash text-to-speech voice models aimed at revolutionizing audio production workflows. The standard Gemini 3.8 Flash TTS model is built for interactive entertainment, game development, and extended narrations that require precise prompt-based voice direction. Meanwhile, the lighter Gemini 3.8 Flash-Lite TTS variant handles automated dubbing, conversational customer agents, and rapid translation pipelines that demand high throughput and efficiency.
Extensive Voice Library and Remixing Features
Both models expand Google's audio toolkit by providing access to more than 2,000 pre-configured vocal profiles spanning over 100 languages, including specialized variants such as Quebec French, Scots English, and Mexican Spanish. Developers can now generate speech through simple text prompts and, in an upcoming update, use a voice remixing module to fine-tune elements like timbre, pitch, speaking pace, and regional accents directly via text instructions without needing separate audio samples.
Strong Performance in Independent Benchmarks
Third-party testing has validated the models' quality. On the Hume AI Voice Design Benchmark, the main Gemini 3.8 Flash TTS model scored 71.4 overall and led in accent modelling with a 60.8 rating. It also topped the Hume AI Overall Quality Index, while the Lite version placed second, surpassing previous Gemini 3.1 Flash TTS results.
Double-blind evaluations in Voice Arena confirmed strong listener preferences for the models in languages including Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish, and Hindi. The systems excel at multi-speaker scenarios, maintaining natural dialogue flow and vocal separation during long conversations, with stable quality even in multi-hour productions like audiobooks or podcasts.
Support for Non-Verbal Audio Cues
Writers can embed production instructions directly into scripts using simple tags such as
Safety and Identity Verification Measures
The voice cloning system requires a 30-second reference audio clip from the original speaker paired with explicit verbal consent. Google verifies acoustic similarity between the reference and new recordings before allowing custom profiles. All generated audio files include imperceptible SynthID watermarks and C2PA metadata embedded in the waveform, enabling easy detection of synthetic content by downstream tools.
Developer and Enterprise Availability
Both models are accessible through Google AI Studio and the Gemini API, with integrations supporting frameworks from Agora, LiveKit, Pipecat, and Vercel. Early partners include Figma, HeyGen, Linguana, Wondercraft, 99.co, and Ollang for applications in media translation and customer service. End users can try the standard TTS model inside Gemini Notebook, while Google Vids incorporates the Lite version. Enterprise users will gain administrative API access in coming updates.