Google Officially Announces Gemini 3.8 Flash TTS, Creating a Studio for Designing Character Voices from Scratch
Google releases Gemini 3.8 Flash TTS
Google officially announces Gemini 3.8 Flash TTS, transforming text-to-speech into a studio for designing character voices from scratch: covering over 100 languages and dialects, supporting dual speaker dialogues, and allowing control over performance details such as laughter, sighs, and breath sounds on a per-line basis.
The model is identified as gemini-3.8-flash-tts, positioned as a flagship creative product, emphasizing studio-level fidelity, performance nuance, regional accents, and long-form multi-turn speech without distortion. The series also includes the high-throughput, low-cost Gemini 3.8 Flash-Lite TTS, aimed at voiceover batches, reading, and real-time voice agents. Both share the same API schema, with switching requiring only a model name change. Official documentation states that the Flash version supports over 130 languages, while the Lite version supports over 100.
The voice system has expanded from about 30 preset tones to over 2000 directly producible voice libraries, including regional variants like Mexican Spanish, Quebec French, and Scottish English. Creators can generate new voices from scratch using natural language to describe character, accent, and voice features; they can also replicate voices using approximately 30 seconds of authorized samples, which default to a persistent voice in the project, with a stateless key provided. The replication process includes consent verification, with outputs marked with SynthID watermarks and C2PA credentials.
Dual dialogues can be rehearsed in a single request: binding preset voices to two speakers, labeling each line with the speaker and style, allowing the model to generate natural responses in turns. Line-level instructions can include cheerful, dialect switching, pauses; embedded tags cover laughter, sighs, gasps, and affirmations like mhm and yeah. The current limit for multi-speaker synthesis is two, and if custom or replicated voices are to be used together, they must be synthesized separately in turns and then stitched together. Audio defaults to 24 kHz mono: complete requests return WAV, while streaming requests return headless PCM.
In evaluation metrics, Flash TTS ranks first with a total score of 71.4 in Hume AI voice design benchmarks and 60.8 in accent modeling; Flash and Lite rank first and second in overall quality index. In blind tests by Voice Arena, both models ranked highly in languages such as Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish, and Hindi. The product will be available on the Gemini API and Google AI Studio starting September 23, with Notebook access available simultaneously, and enterprise interfaces to follow with Gemini Enterprise.
In market mechanics, this transforms the production of audio content from studio voiceovers and editing pipelines into a pay-per-use generative interface. Buyers include audiobook, podcast, game dialogue, localization voiceover, and voice agent teams, purchasing directive performance control rather than just another reading tone. Google sells high-fidelity creative premiums with Flash and charges for bulk duration with Lite. Funding shifts from studio labor hours and tone licensing to API inference and storable voice assets. Beneficiaries are developers already creating content tools within the Gemini ecosystem; those under pressure are traditional TTS vendors selling preset tones and mid-tier studios still relying on multi-track stitching for dual podcasts.
Supplementary data: Long-form stability is stated to maintain tone, volume, and room feel without drifting over several minutes of narration; some materials claim rhythm and tone can sustain for hours of continuous audio. Preset studio voices still retain about 30 premium names, such as Kore and Puck, while the expanded library can be searched through the voices interface by language, gender, pitch, and scene. Custom and replicated voices can be stored in projects for about a year by default.
Source: Public Information
ABAB AI Insight
Google has transformed TTS from a "reading model" into a "directing interface," following the same path as its image and video generation: first lowering the entry cost with preset assets, then handing control to natural language prompts. The Gemini 2.5/3.1 preview versions only supported single and dual-person reading with limited preset voices, while 3.8 integrates voice design, replication, line-level performance, and regional dialects all at once. This aligns with DeepMind's consistent strategy of unified multimodal models: voice is no longer an independent product line like WaveNet, but one output modality on the Gemini generative interface.
The capital path turns voice lines into storable, reusable, and watermarked digital assets. The 30-second replication with consent verification indicates that Google aims to achieve two things: allowing creators to have "their own voice" while shifting the legal responsibility for celebrity voice misuse to the credential layer. The separation of Flash and Lite is a typical cloud pricing model: the creative side uses high fidelity to lock in game and audiobook production studios, while the agent side uses low latency to capture real-time customer service and reading buttons. The one-year storage of voice_id creates switching costs for projects, thus transforming inference fees from one-time listening to continuous subscription consumption.
Comparable examples include ElevenLabs turning voice cloning into creator infrastructure, and OpenAI using GPT-Live for real-time dialogue while reserving precise reading for another TTS. Google is targeting "directable content production," rather than phone bots competing for speech. The industry phase has shifted from expanding language varieties to controlling performance: whoever can arrange dual dialogues, punchlines, and dialects in a single request will dominate the intermediate processes of podcasting and game localization. If editors like Adobe and Descript integrate, the editing timeline will shift from "waveform matching" to "line annotation."
Structural judgment belongs to technological substitution. The mechanism is: the scarcity of in-studio performances comes from recording studios, actor schedules, and post-production alignment; when laughter, breath sounds, and dual turns can all be written into tags, scarcity shifts to the quality of prompts and voice line copyrights. Watermarks and C2PA externalize detection costs, allowing generators to exchange compliance for enterprise procurement. The result is not the disappearance of the voiceover industry, but a shift in pricing from hourly labor to character-based inference, and a transfer of directing power from mixing consoles to API fields.
ABAB News · Cognitive Laws
Performances that can be annotated will be consumed by interfaces.
Preset tones attract customers, while persistent voice lines retain them.
Watermarks are not courtesy; they format responsibility.