Build2 publishers2 min readPublished
Gemini 3.8's cheap TTS tier still takes line-by-line stage directions
Google shipped two text-to-speech models with the same per-line performance control and put voice creation on only one of them, so the tier decision depends on whether a project has to invent a voice at all.
The Engineer · Build desk
What happened
- Google added two text-to-speech models to the Gemini family, Flash TTS for deep creative direction and character design and Flash-Lite TTS for high-volume, cost-efficient scale.
- Only Flash TTS is named for generative voice design, building bespoke voices from scratch by prompting for role, accent and voice characteristics across more than 100 languages and dialects.
- Voice replication recreates a vocal profile from a 30-second sample, with built-in consent verification, SynthID watermarking and C2PA credentials attached to the output.
- Google places Flash TTS first and Flash-Lite second on Hume AI's Overall Quality Index, the only head-to-head ranking it cites for the cheaper model.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision A dubbing queue or voice agent with a fixed cast can render on the cheaper model and keep its acting cues. Expressiveness is not what separates the tiers; the tier choice turns on whether the workload invents voices.
- constraint Google did not publish a price for either model, so nobody outside can compute the monthly volume at which routing to Flash-Lite repays the engineering to split the pipeline.
- exposure Thirty seconds is short enough that almost any recording a studio already holds is a usable sample. That shifts the legal risk on talent deals from the session booking to the consent verification step.
- capability Two-speaker staging from one script and hours of continuous audio with minimal drift remove the stitching and re-take passes that dialogue and audiobook pipelines currently budget for.
Both models take direction. Google's post says both TTS models give the user precise control over how each line is delivered [4], and it describes Flash-Lite with fine-grained control over tone, pacing and expressive nuance [3]. The performance section, which covers both, lists the same nonverbal tag syntax: `<laughs>`, `<sigh>`, `<gasp>`, plus active-listening interjections like `|mhm|` and `|yeah|` [10].
The catalogue sits on the Flash side. Google describes the voice studio as scaling up from 30 original voices to 2,000-plus production-ready voices, including regional varieties like Mexican Spanish, Quebec French and Scots English [6][7]. That is roughly 66 times the number of named voices [22].
Which raises the handoff question for anyone building a two-tier pipeline. Google says custom voices can be saved and managed to hold consistent performance with minimal drift across ongoing projects [9], but the post does not say whether a voice designed in the Flash studio can then be rendered by Flash-Lite. Without that, design-on-Flash and render-on-Flash-Lite is an assumption.
The quality evidence is third-party and thin on method. Google reports Flash TTS first overall on Hume AI's Voice Design Benchmark with 71.4, and leading accent modeling at 60.8 [11]. Blind human preference evaluations on Voice Arena put Flash and Flash-Lite in top positions in key global languages [12]. The 71.4 comes with no stated scale and no named field of competitors. It is a claim about someone else's prompt set. For it to transfer, the benchmark's briefs have to resemble yours: how many ask for an invented character, and how many just need a calm agent reading a refund policy.
The internal comparison is more legible. Google says the new models show major improvements on long-form content and dual-speaker screenplay control compared with Gemini 3.1 Flash TTS [13]. If you already run 3.1 Flash TTS on audiobook or dialogue work, you can measure that delta on your own scripts.
The two models follow 3.5 Live Translate, 3.5 Transcribe, 3.8 Live and 3.8 Live Extended Thinking in the Gemini Audio family [14], and Google says they improve user experiences in products including Gemini Notebook and Google Vids [15]. Voice remixing, which fine-tunes timbre, pitch, pace and accent from prompts such as "add subtle Southern US accent" or "soften the delivery", is listed as coming soon [16].
What to watch
- Whether Voice Arena's public leaderboard reproduces the ordering Google reports once third parties run the same languages.
- Whether SynthID detection for replicated-voice output is exposed to the people buying the audio, or only inside Google.
- Whether the voice designer reaches end users in Google Vids and Gemini Notebook. That would change who the consent check is protecting.