Google has introduced two new text-to-speech models, Gemini 3.8 Flash TTS and Flash-Lite TTS, supporting over 100 languages. Flash TTS can create new voices from text descriptions, and both models allow users to add stage directions to individual lines and generate two-voice dialogue from a single script. A voice cloning feature can build a voice profile from a 30-second sample.
SynthePulse Insight · AI deep readingMembers
Version 1 · 1 source
Google released Gemini 3.8 Flash TTS and Flash-Lite TTS, supporting over 100 languages, voice generation from text descriptions, and 30-second audio cloning, with clear token-based pricing. But Voice Remixing is not yet live, regional endpoints are not listed, and hands-on audio still shows high-frequency squealing—there is a clear gap between expanding capability and engineering maturity.
Google released two text-to-speech models, Gemini 3.8 Flash TTS and Flash-Lite TTS, both supporting over 100 languages. This is the most basic factual layer of the launch: not an iteration of a single model, but two product lines with different positioning rolled out at the same time.
By Google's positioning, Flash TTS targets creative scenarios such as game characters, audiobooks, and podcasts, while Flash-Lite TTS focuses on low-cost, large-scale voice generation for dubbing, audio content, and voice agents. This division determines the differences in later features and pricing: the creative type emphasizes controllability and voice design, while the scale type emphasizes unit cost.
The two models share a set of baseline capabilities: both can add stage directions to each line, support two-person dialogue mode, and include nonverbal sounds such as laughter and sighs. In other words, script-level expressive control is the common foundation of both product lines, not something unique to Flash TTS.
The core selling point of Flash TTS is "designing a voice from scratch": users can use text prompts to define a character's role positioning, accent, and voice characteristics, covering multiple languages and dialects. This is equivalent to turning voice production from "picking an existing voice" into "describing a voice in natural language."
For users who do not want to start from scratch, Google provides a voice library of over 2,000 preset voices, including regional variants such as Mexican Spanish, Quebec French, and Scottish English. The preset library and text-generated voices form two parallel entry points: the former lowers the barrier to entry, while the latter provides room for differentiation.
Another entry point is voice cloning: a voice profile can be built from a 30-second audio sample. But this capability comes with clear compliance constraints—the person being cloned must record a spoken consent statement, and the voice in that recording must match the sample. This turns "authorization" from a procedural declaration into verifiable audio evidence.
At the traceability level, according to Google, every piece of audio generated by Gemini audio models carries an inaudible SynthID watermark to help detect AI-generated speech. Note that this is Google's own statement, and the source does not provide an independent assessment of the watermark's actual detection effectiveness.
The factual basis of this article comes from THE DECODER's report (article_id 18015), which cites official Google information and includes the author's own testing. Among these, the model launch, language coverage, preset voice library size, cloning consent constraints, stage directions and two-person dialogue, distribution channels, developer platform integration, missing regional endpoints, Voice Remixing not being live, and the pricing table are confirmed facts; SynthID watermark effectiveness, minimal speaker drift, leading on the Hume AI benchmark, differences in data usage between free and paid tiers, and the hourly cost converted from tokens are all labeled as Google's own statements, with no third-party verification in the source; the high-frequency squealing and the voice change at the end are observations from the source author's own testing, with a limited sample size, and cannot be generalized into a universal conclusion.
Primary report
Primary source