Google Announces Gemini 3.8 Flash TTS and Flash-Lite TTS Models
Originally published in Turkish. Read the original →
Google has made the Gemini 3.8 Flash TTS and Flash-Lite TTS models available in the API and AI Studio. With over 2,000 voices, voice cloning and scene direction, a new era in AI speech generation has begun.
Google has taken its multimodal AI vision to a new dimension in speech and audio technology. Officially announced on September 23, 2026, Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS were introduced as the most expressive text-to-speech models the company has built to date. Developed by the Gemini Audio team within Google DeepMind, these two new models open a new chapter across a wide range of fields, from interactive games to audiobooks and from call center solutions to global dubbing and media production. Available to developers on the Gemini API and Google AI Studio from the moment they were announced, the models noticeably raise the bar for voice AI.
Two Developer-Focused Models: Flash TTS and Flash-Lite TTS
To balance creativity against operational cost in speech generation, Google designed a model lineup with two different focal points. Both variants are built on the Gemini 3 Pro architecture and support up to 8 thousand tokens of text input and up to 64 thousand tokens of audio output.
Gemini 3.8 Flash TTS was built for projects that require artistic direction and sophisticated character design. It stands out in areas where narrative power is critical, such as dynamic dialogue in video games, multi-layered podcast series and immersive, atmospheric audiobooks. Content creators can shape a voice's tone, breathing, pauses within a scene, emphasis and emotional ups and downs in fine detail.
Gemini 3.8 Flash-Lite TTS, on the other hand, is optimized for enterprise projects that target high-volume operations and budget efficiency. It is an ideal fit especially for global video dubbing, fast content localization and voice dialogue agents that must handle thousands of concurrent sessions. Running with low latency, the model offers precise control over speaking pace, pitch and natural expressive nuance, bringing studio-quality voice at economical prices.
Natural-Language Voice Design and a Rich Voice Library
One of the most striking aspects of the new models is that they make designing an original voice identity from scratch practical. Without wrestling with complex acoustic settings, developers can describe the voice character they want simply through prompts written in natural language. By specifying age, gender, professional demeanor, accent and momentary emotional state, a unique voice profile can be generated in seconds.
This flexibility also shows in global language support. The system supports more than 100 languages and regional dialects. Going beyond the major languages, local variants such as Mexican Spanish, Quebec French and Scottish English can be voiced naturally. With this launch, Google turned its collection, previously limited to 30 voices, into a huge library of more than 2,000 production-ready voices. Users can use the custom voices they create consistently across their projects, and with a voice mixing feature coming soon, the timbre and tempo of existing voices will also be flexibly transformable.
Natural Scene Direction and Dual-Speaker Staging
Speaker drift in long recordings, one of the biggest handicaps of traditional TTS systems, is solved with the Gemini 3.8 models. According to the shared data, even in uninterrupted audio generation lasting hours, voice quality, speaking rhythm and character timbre are preserved without degradation.
The models also support dual-speaker scene direction natively. Developers can create lively dialogues in which two different characters talk to each other from a single script. The characters speak without talking over each other, with realistic pauses and natural turn-taking. Reactions added to the script to give the scene depth, such as laughing, sighing and surprise, as well as acknowledgment sounds, are performed by the AI with genuinely human naturalness.
Safety Standards, Consent Verification and SynthID
To prevent the abuse that voice cloning technology can enable, Google strengthened its safety layers. The voice replication capability, which works with only a 30-second reference recording, is protected by a strict verbal consent verification mechanism. To clone a voice, a spoken consent recording matching the identity of the voice's owner must be submitted to the system. Because of biometric data regulations, this feature is subject to legal restrictions on AI Studio in Illinois, Texas, the EU, the United Kingdom, Switzerland and India.
Every generated audio file is marked with SynthID watermarking technology developed by Google DeepMind. Embedded directly in the waveform and inaudible to the human ear, this acoustic signature makes it possible to verify that the content was generated by AI. C2PA authentication, which provides global content transparency, is also part of the infrastructure.
Benchmark Results and the Developer Ecosystem
With their performance in independent test environments, Gemini 3.8 Flash TTS and Flash-Lite TTS have taken the top spot in the audio space. Flash TTS ranked first overall with a score of 71.4 in the Voice Design Benchmark conducted by Hume AI, and also took the lead in accent modeling with a score of 60.8. The two models share the top two places in the Hume AI Overall Quality Index, and in blind tests on Voice Arena they were users' clear favorites in languages such as Japanese, Brazilian Portuguese, Vietnamese, Arabic, Mexican Spanish and Hindi.
A broad developer ecosystem also mobilized with the launch. LiveKit, Pipecat, Agora and Vercel added support for the new voice models right away, while global brands such as Figma, HeyGen, Linguana, Wondercraft and Ollang began using this infrastructure in multilingual dubbing and voice assistant solutions. The new audio workspace in Google AI Studio lets developers test voice identities and direct scripts live without writing code.
