Google launches Gemini 3.8 Flash TTS and Flash-Lite TTS models

By Splitfeed EditorialLast updated Sep 26, 2026, 19:17 UTC

On September 23, 2026, Google released two new text-to-speech models, Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, through the Gemini API and Google AI Studio. The models support more than 100 languages, offer over 2,000 voices, allow custom voice design from text prompts, and include voice replication. Every clip carries SynthID watermarking, an imperceptible identifier for AI-generated audio. Google also announced the models would roll out to Google Vids and Gemini Enterprise in the near future.

The inclusion of voice replication is the detail that gives the release its sharpest edge. Voice cloning has historically been treated cautiously by major platforms, so its open availability here marks a shift in what large providers are willing to ship commercially. SynthID watermarking is Google's stated answer to misuse concerns, though watermarking is a deterrent rather than a hard barrier. The Flash and Flash-Lite tier structure also signals a push for cost efficiency, with simonw noting that most individual experiments cost less than a cent.

The Pro-AI camp is largely enthusiastic. ttul tested the model as a replacement for ElevenLabs' v3 in a live application and found the output "extremely expressive" and on par with the competitor. simonw highlighted the low cost. Google's own communications framed SynthID as a transparency feature that protects creators. Critics are less convinced. yipinwong argued Google is "spreading too thin" across too many simultaneous AI products without coherent focus, while rcr-anti pointed to inconsistent availability across Google's consumer, prosumer, and cloud platforms. nater5000 found nothing particularly impressive, and Leon Skum questioned whether any of these products generates meaningful revenue outside government contracts. The middle ground camp views the release as evidence that voice cloning has become commoditized, with simonw observing that Google's willingness to ship it reflects how widely available the capability already is elsewhere. Even so, several listeners find the output unconvincing: burkaman said results fall short of prompt accuracy, m3kw9 described the voices as over-expressive in an artificial way, and AyanamiKaine suggested that while many listeners may not notice, something present in real human speech remains absent.

The practical test will come as Google rolls out the models to Google Vids and Gemini Enterprise. Adoption rates in those products, alongside any policy response to the voice replication feature, will clarify whether Google's open approach to voice cloning becomes an industry norm or prompts regulatory pushback.