Google DeepMind Introduces Gemini 3.8 Flash TTS and Flash-Lite TTS for Prompt-Directed Voice Generation

Google DeepMind has introduced Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, adding prompt-directed speech generation to the Gemini 3.8 audio lineup. Gemini 3.8 Flash TTS is positioned for controllable character and creative voice design, including new voices described in natural language, and the launch lands as Google's third audio-model push in less than 30 days. The Gemini 3.8 Audio model card lists the two text-to-speech models alongside Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, and says the TTS variants take text inputs up to 8K tokens and output audio with a 64K-token output budget. The card says Gemini 3.8 Audio is based on Gemini 3 Pro, carries a January 2025 knowledge cutoff, and is distributed through Gemini API and Google AI Studio; Gemini 3.8 Flash TTS is also listed for Gemini Notebook, while Gemini 3.8 Flash-Lite TTS is also listed for Google Vids. It also flags general foundation-model limits such as hallucinations, jailbreak risks under continuing mitigation, and occasional slowness or timeout issues. The launch extends a rapid audio sequence. On August 26, Google's Gemini API release notes recorded Gemini 3.5 Transcribe and Gemini 3.5 Transcribe Live, with 85+ languages, speaker diarization, word-level timestamps and custom vocabulary biasing of up to 1,000 terms. On September 15, Google announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, saying Live supports 97+ languages, asynchronous function calling and visual context, while Transcribe had measured average Word Error Rate of 4.0% for streaming and 2.6% for non-streaming. The new TTS models fill the other side of that stack: not listening or real-time dialogue, but generating finished speech for apps, videos, agents and character-driven content. Google's baseline for this push was Gemini 3.1 Flash TTS, released in April as an expressive speech model with audio tags, support for more than 70 languages and SynthID watermarking. Google reported at launch that it reached 1,211 Elo on the Artificial Analysis TTS leaderboard and highlighted natural-language control over style, pace and delivery. The DeepMind speech-generation page frames Gemini TTS uses around immersive storytelling, multi-speaker generation, AI avatars, customer-support agents, education and localized video. Developer reaction is likely to split along those same lines. The upside is the move from choosing a voice to directing a performance: more precise personas for podcast-style dialogue, training simulations, video narration and voice agents. The pressure test is production consistency. Users of earlier Gemini TTS workflows have reported issues such as micro-cuts and pitch jumps in some generations, and broader discussion is focused on whether the 3.8 TTS models improve naturalness, long-form reliability and multilingual consistency. The new models therefore arrive with a clear opportunity and a clear burden of proof: stable voice identity and production-grade audio quality will matter as much as natural-language character design.