Google Launches Gemini 3.8 Flash TTS and Flash-Lite TTS with Voice Creation from Text Prompts
Google DeepMind has released Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, text-to-speech models that generate custom voices from natural language prompts and support line-by-line performance direction. The models top Hume AI's Voice Design Benchmark at 71.4 and claim first and second place on its Overall Quality Index.
Gemini 3.8 Flash TTS at a glance
Google DeepMind released two new text-to-speech models on September 23, 2026: Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS. The models are available now in Google AI Studio, the Gemini API, Gemini Enterprise, Gemini Notebook, and Google Vids.
Gemini 3.8 Flash TTS is positioned for creative direction and character design, letting users generate original voices from natural-language prompts across more than 100 languages and dialects. Gemini 3.8 Flash-Lite TTS targets high-volume use cases like dubbing and voice agents, optimized for cost efficiency at scale. Pricing not yet disclosed for either model.
Benchmark Claims
According to Google, Gemini 3.8 Flash TTS holds the #1 overall spot on Hume AI's Voice Design Benchmark with a score of 71.4, and leads in accent modeling with a score of 60.8. The company also claims Gemini 3.8 Flash TTS and Flash-Lite TTS take the #1 and #2 spots, respectively, on Hume AI's Overall Quality Index. In blind human preference evaluations on Voice Arena, Google states the models rank at the top among competitors for Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish, and Hindi. None of these benchmark claims have been independently verified.
What's New
The models expand voice options from 30 preset voices to what Google describes as an "infinite library" through generative voice design — users describe a voice ("a fire-breathing dragon" or "a charismatic narrator with a regional cadence") and the model generates it. Google AI Studio also ships with over 2,000 production-ready voices covering regional variants like Mexican Spanish, Quebec French, and Scots English.
Voice replication lets users recreate a vocal profile from a 30-second audio sample, gated by a consent-verification step requiring a matching verbal consent recording from the voice owner. Other features include line-by-line performance direction with acting cues, long-form generation aimed at minimizing speaker drift across hours of audio, native two-speaker scene staging for dialogue, and scripted non-verbal cues like laughs, sighs, and backchanneling interjections.
A "voice remixing" feature — letting users adjust timbre, pitch, pace, and accent of existing library voices via prompts — is listed as coming soon rather than available at launch.
Every generated audio clip is watermarked with SynthID, and outputs carry C2PA credentials for provenance tracking. Google says this is intended to keep AI-generated speech detectable and reduce misinformation risk.
The release follows other entries in Google's Gemini Audio lineup, including 3.5 Live Translate, 3.5 Transcribe, 3.8 Live, and 3.8 Live Extended Thinking, and is compared favorably by Google against its predecessor, Gemini 3.1 Flash TTS, particularly on long-form content and dual-speaker screenplay control.
What this means
Google is pushing TTS beyond fixed-voice presets toward a prompt-driven voice design workflow, competing directly with players like ElevenLabs in synthetic voice generation. The consent-verification requirement for voice cloning and mandatory SynthID watermarking signal Google is trying to get ahead of misuse concerns before regulators force the issue. The lack of published pricing makes it hard to judge whether Flash-Lite TTS actually delivers on its cost-efficiency promise for high-volume applications like dubbing and voice agents — that comparison will matter more than benchmark rankings once developers start building at scale.
Related Articles
Underdog releases Woof, a 31M-parameter open text-to-speech model that runs in a browser
Underdog's new text-to-speech model Woof has 31 million parameters and weighs 31.7 MB. The company says it runs on most devices without an internet connection, and the weights are open to download. Underdog demonstrated it with a browser game posted on November 5, 2026.
Nous Research releases Arete-27B, a Qwen3.8-27B fine-tune that cuts propaganda rate from 22% to 0.9%
Nous Research fine-tuned Alibaba's Qwen3.8-27B on about 53 million tokens to remove Chinese state propaganda from its answers. According to Nous, the resulting Arete-27B model drops the propaganda rate from 22 percent to 0.9 percent while keeping general capabilities intact. Weights, test set and training data are public under Apache 2.0.
PrunaAI releases P-H3-Image: 8 MP image generation and editing in 8 steps, built on MiniMax-H3
PrunaAI has released P-H3-Image, an image model based on MiniMax-H3 that renders 8-megapixel images in 8 steps with no classifier-free guidance. It handles text-to-image and editing with up to 8 input images, and bf16 inference needs about 101 GB of GPU memory.
OpenCSG Agentic-27B: 27.36B-Parameter Tool-Use Model Runs on One A800, Commercial Use Needs Permission
OpenCSG has released Agentic-27B, a 27.36B-parameter model for multi-step tool use, built on the dense Qwen3.8-27B. The company reports it runs on a single 80GB NVIDIA A800, but commercial use requires authorization under the OpenCSG license, and all benchmark results are OpenCSG's own.
Comments
Loading...