Google Releases Gemini 3.8 Flash TTS Voice Models: Direct Voice Acting Line by Line Like a Director

Google has added two new text-to-speech (TTS) models to the Gemini family: Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS.

Previously, using TTS generally followed the same routine: choose one voice out of several dozen presets, feed in text, and receive an audio track read with reasonable fluency. If you wanted the voice to sound "a bit more nervous," "deliver this line with a smile," or have "the other person chime in with an 'uh-huh' on the side," it was either impossible or required manual editing in post-production.

With this release, Google is turning the process into a "voiceover studio": voices can be designed from a single sentence or replicated from a 30-second recording; line-by-line acting can be directed just like writing a screenplay; and two-character dialogue, interruptions, sighs, and laughter can all be generated from a single script.

Two Models: One for "Great Acting", One for "High Volume and Low Cost"

The division of labor between the two models is straightforward:

In short: choose Flash for characters and dramatic scenes; choose Flash-Lite when generating tens of thousands of voiceover lines or customer service dialogues per day. Google did not include pricing details for either model in its launch post.

Notably, Flash-Lite is not a "watered-down" version. In several benchmark evaluations discussed below, its scores run nearly neck-and-neck with Flash, and in blind listening preference tests across three languages, including English, it even ranks ahead of Flash.

These two models belong to the expanding Gemini Audio family, joining earlier releases such as 3.5 Live Translate (real-time translation), 3.5 Transcribe (transcription), 3.8 Live, and 3.8 Live Extended Thinking (real-time voice conversation).

Describe Any Voice You Want in a Single Sentence

Traditional TTS only allowed you to "select" a voice; this time, Flash lets you "create" one. Google calls this Voice Design: write out the character, accent, and vocal traits in natural language, and the model synthesizes a new voice from scratch, covering 100+ languages and dialects.

Google shared several sample clips generated from concise prompts:

• "High-energy DJ from Melbourne": Produces an Australian-accented radio host speaking with rapid delivery and upbeat enthusiasm: "All right, Melbourne... Heaps of great music on the way for you", even capturing local Australian slang like "heaps".
• "Super high-pitched, monotone robot": The entire line "alert alert system status report major malfunction..." plays without inflection, capturing the tone of vintage sci-fi computer broadcasts.
• "Japanese dragon": A dragon disturbed from its slumber refers to itself as "washi" (an archaic first-person pronoun used by older figures) and interrogates the intruder in Japanese: 「わしの眠りを侮ったのか、それとも……所詮は好奇心か」 ("Did you make light of my slumber, or... was it mere curiosity after all?").

The most intuitive showcase is a game development demo. Traditional character customization in gaming rarely goes beyond visual appearance, offering at most a few voice presets. In the demo, the developer enters the following into the "Voice Design" prompt field:

Deep gravely voice that shows he has been through many battles. Fast and booming.

Clicking "Create Voice" produces, within seconds, a knight voicing the line: "You shall not pass these gates, demon, for I have battled worse than thee". Changing the prompt to a high-pitched princess with royal poise immediately converts the exact same dialogue into a princess's voice. Players can now customize character voices just as they customize faces.

Several complementary features accompany Voice Design:

• Expanded voice library: Expanded from 30 preset voices to 2,000+ ready-to-use voices, including regional varieties like Mexican Spanish, Quebec French, and Scottish English.
• Save and reuse: Designed voices can be saved and cataloged. This is crucial for episodic content, where the protagonist's voice in episode 1 and episode 50 must match; Google states it achieves minimal drift.
• Voice fine-tuning (coming soon): Select an existing voice from the library, then adjust timbre, pitch, pace, and accent using prompts, such as "add a slight Southern American accent" or "make it sound softer".

Replicating Your Voice from an Audio Sample

Another way to acquire a voice is voice replication: take a sample of your own voice, or a voice you have the rights to use, and generate a digital version capable of reading arbitrary text. The official blog mentions needing only 30 seconds of audio; the Google AI Studio demo interface suggests recording approximately 20 seconds of natural speech, with the tip: "tell a story, don't read in a flat monotone".

The workflow demonstrated in the video proceeds as follows:

1. Record a sample: Read aloud for about 20 seconds using the provided prompt text (a short narrative about a lighthouse keeper gazing across the ocean horizon).
2. Verify identity: Read a consent statement aloud: "I am the owner of this voice and I consent to Google using this voice to create a synthetic voice model." The system cross-checks this consent recording against the previous sample to verify they originate from the same speaker.
3. Generate and save: Once analysis is complete, the replicated voice says, "Hi there! I am your newly replicated voice". You assign it a name, save it to your voice library, and call it via the Gemini API.

Step 2 represents what Google calls consent verification: the voice owner must verbally record consent that matches the reference audio before generation is permitted. This prevents unauthorized cloning using another person's recordings.

Additionally, voice replication in Google AI Studio is currently unavailable in Illinois, Texas, the European Economic Area, the United Kingdom, Switzerland, and India. Developers located in these jurisdictions cannot use the feature at this time.

Directing Every Line Like Writing a Screenplay

Once a voice is selected, both models support line-by-line directing: you can write stage directions directly into the script instructing whether a line should be muttered, reassuring, or whispered; or you can leave directions out and allow Gemini to determine delivery from the dialogue context.

Based on Google's demonstration, scripts utilize three types of notation:

• Square brackets [ ]: Performance directions. Placed before a line to indicate tone and emotional state, such as [whispering], [reassuringly], or extended descriptive phrases.
• Angle brackets < >: Non-verbal sounds. Embedded directly within dialogue, such as <sigh>, <laughs>, <gasp>, or <breath>.
• Vertical bars | |: Backchanneling and interruptions from the other speaker. Markers like |mhm| or |yeah| deliver active listener feedback (which Google refers to as backchanneling, capturing conversational listener reactions like "uh-huh" or "right") while also pacing comedic timing and reaction beats.

Consider this single-speaker customer service example, where the agent moves through four distinct emotional states across four lines:

Speaker 1: [muttering]
tu...tu...tu....hang on a second, let me find that order status <sigh>

Speaker 1: [reassuringly]
yeah....this order never really got delivered. Sorry for that. Let me update and reschedule this order....

Speaker 1: [whispering]
give me one moment...

Speaker 1: [warm][neutral][prosody="60%"]
I just updated the status and rescheduled the delivery.

The delivery mirrors a real customer service representative: muttering and sighing while searching, softening into an apologetic tone upon discovering the problem, whispering "one moment" while updating the record, and finishing with a composed, warm tone to confirm resolution. The final line also introduces a parameterized tag, [prosody="60%"]; while the blog does not explain its exact mechanics, it contextually regulates intonation and rhythmic range.

Next, consider a two-character horror scene employing all three notations simultaneously, as two individuals in a dark forest realize something is stalking them:

Speaker 1: [high-pitched and frantic, trying to rationalize while escalating into pure, unadulterated terror]
Maybe it's just a deer? |That's not| <breath> People say |That's not a deer.| it's always deer. <nervous laugh> But it's... it's staying right on the edge of the light. |Stop.| It's following us. <gasp> Look, right there |Where?| by the pine!

Speaker 2: [strangled and horrified, voice cracking as the realization of a threat sinks in]
I don't see... |There!| <gasp> Oh my god, what is |What is that?| that? <breath> It's too big. That's not |That's not| an animal. It's just standing there.

Two implementation details here are critical.

First, stage directions describe "emotional arcs", not merely static emotions. "High-pitched and frantic, trying to rationalize while escalating into pure, unadulterated terror" specifies a dynamic progression. Later directions include "hushed and urgent, teeth clenched to maintain composure" and "sobbing and breaking down, volume and pace escalating as threat nears".

Second, interruptions marked inside vertical bars overlap with the primary dialogue. When Speaker 1 says "Maybe it's just a deer?", Speaker 2 is already cutting in with "That's not". This mirrors how real people argue, panic, and interrupt, an effect impossible with conventional TTS engines that process dialogue in alternating, sequential turns.

These capabilities rest on several underlying foundations:

• Native multi-speaker scenes: A single script directly coordinates multi-turn dialogue between two speakers, maintaining clear vocal distinction, natural turn-taking, and conversational overlap. This suits podcast dialogues and audio dramas.
• Long-form audio generation: Generates hours of continuous audio while maintaining consistent quality, cadence, and timbre, with minimal speaker drift. In audiobook production, this eliminates the risk of a character's voice shifting ten chapters in.
• Non-verbal vocalizations and backchanneling: The angle bracket and vertical bar notations highlighted above, which Google emphasizes for shaping comedic timing and conversational reaction points.

Google also provided two comprehensive "director" demos.

A seasoned detective monologue. A demo application named Monologue breaks a monologue into 16 segments, each headed by explicit performance instructions, such as "smug and relaxed, leaning back on a metal chair, slow and amused Brooklyn accent," "shouting at top volume, veins bulging with rage over betrayal," and "heartbroken, voice cracking and wavering, fighting back tears." The script incorporates <chuckle>, <short pause>, <breath>, and phonetic notation /kəˈnɑːrsi/ to guide the pronunciation of "Canarsie" and prevent misreadings. The result is a complete emotional arc as an aging detective shifts from arrogant to panicked, then broken, before finally disclosing the clue.

Script table reads. Upload a script, assign voices to each character (such as a "warm, resonant baritone" or a "gentle, empathetic mezzo-soprano"), and the scene is performed in full. Its guiding premise: "Change only the performance, without altering a single word of dialogue." In the demo, a creator alters a stage direction to "complaining but conversational," regenerates the line, and hears the identical text delivered under a completely new emotional interpretation. For writers and directors, this offers an on-demand cast of voice actors for iterative rehearsal.

To try this workflow, scripts can be structured following this template: define the speaker, place stage directions (including emotional progression) in square brackets, embed non-verbal sounds in angle brackets, and place overlapping interruptions from the other speaker within vertical bars.

Speaker 1: [Performance direction: emotion, volume, pace, and how emotion evolves across the line]
Line... <sigh> Line... |Interruption from second speaker| Line...

Speaker 2: [Performance direction]
Line... <laughs> Line... |mhm| Line...

Benchmark Performance: Where It Leads, Where It Falls Behind

Google cited two external benchmarks: leaderboards from Hume AI and blind preference voting from Voice Arena. All three tables originate from Google's launch post, with higher numbers indicating better performance.

Table 1: Hume AI Voice Design Leaderboard, evaluating the task of "creating voices from text descriptions," covering three competitors: Gemini 3.8 Flash TTS, ElevenLabs Voice Design v3, and Inworld Voice Design.

Gemini ranks first in overall score, though leading ElevenLabs by only 0.6 points. The decisive gap appears in Accents: 60.8 versus 45.4, a margin of over 15 points. This aligns with earlier examples like the "Melbourne DJ" and "Scottish English," showing that its strongest advantage lies in rendering localized dialects. Conversely, Gemini finishes last among the three in "Vocal Traits".

Table 2: Hume AI Quality Leaderboard, evaluating comprehensive TTS performance. The overall score is defined as "Reliability × Expressiveness," capturing whether speech is both stable and emotive; Style Tag Control tests how faithfully the model adheres to individual prompt tags.

Three takeaways emerge from this data:

• Substantial generation-over-generation progress: Overall score increased from 0.783 on 3.1 Flash to 0.920, and multi-speaker capabilities rose from 3.60 to 4.14, reflecting what Google describes as "significantly improved two-character screenplay control."
• Flash-Lite nearly matches Flash: The gap across every metric remains within 0.07 points. For cost-sensitive workflows, Flash-Lite incurs minimal quality degradation.
• ElevenLabs remains stronger in "Humanlike Variation": This metric evaluates whether conversational inflections sound authentic across multi-turn exchanges, penalizing both robotic monotony and exaggerated acting. ElevenLabs v3 achieved a perfect 5.00 compared to Gemini's 4.58; however, on the combined "Reliability × Expressiveness" metric, ElevenLabs v3 scored 0.706.

Table 3: Voice Arena Language Blind Listening Leaderboard. Human evaluators voted without knowing model identities; scores represent relative rankings derived from aggregated votes. Competing external models included ElevenLabs v3, Cartesia Sonic 3.6, and OpenAI gpt-4o-mini-tts, with the table below showing the top-performing external competitor for each language; OpenAI ranked lowest across all seven languages (Arabic evaluated as Modern Standard Arabic, MSA; Cartesia had no scores recorded for Japanese, Vietnamese, or Arabic).

Across all seven languages, the 3.8 family took first place, with the two models splitting victories: Flash led in Japanese, Arabic, Hindi, and Mexican Spanish, while the lower-cost Flash-Lite finished ahead in English, Brazilian Portuguese, and Vietnamese. In English, Flash placed third overall, trailing Cartesia by 7 points. In Hindi, Flash topped Cartesia by only 2 points, essentially a tie, while Flash-Lite (1076) trailed the older 3.1 Flash (1086).

The most pronounced advantage appeared in Japanese: 1,232 points, an 80-point lead over second place, consistent with feedback from Japanese developers.

It should be noted that these benchmarks were selected by Google for its launch blog, meaning comparison targets and metrics were chosen by the company. While the performance leads are substantial, examining the full tables also reveals areas where ElevenLabs leads in "Humanlike Variation" and "Vocal Traits".

What Early Adopting Companies Value Most

Google highlighted several partner integrations. Platforms such as Agora, LiveKit, Pipecat, and Vercel have integrated the models via the Gemini API, enabling developers to build voice applications directly. Additional integrations are underway with Figma, HeyGen, Linguana, Wondercraft, 99.co, and Ollang, focusing primarily on global dubbing, localized accent adaptation, and voice Agents.

Partner testimonials highlight several common themes:

• Japanese accuracy: The developer of Katsuyo noted that other models tend to either "over-act" or make errors with pitch accent and kanji pronunciation; 3.8 Flash was chosen because it delivered the most accurate and natural output across both emotional dialogue and daily speech. This corresponds with its first-place ranking in Voice Arena's Japanese evaluation.
• Regional accents: 99 Group (99.co) uses the model for real-time AI telephone conversations, specifically citing native support across localized languages and accents ranging from Singaporean English (Singlish) to Indonesian.
• Vocal consistency: The most frequently referenced quality. AI video platform HeyGen noted that its users have created over 20 million custom voices (cumulative across HeyGen's platform) and that Voice Design delivers greater voice consistency; Transforms.AI, which builds characters and lore for the Simia app, stated that stable voices are essential for sustaining deeper narrative immersion; YouTube localization service Linguana emphasized the voice stability required for high-volume production pipelines.
• Audiobooks: AI audiobook company Spoken analyzes manuscripts for tone, accent, and character dynamics before passing them to 3.8 Flash for delivery, citing its "expressive vocal range" and "large context window."
• Scaled content generation: Podcast and audio advertising platform Wondercraft selected Flash-Lite, citing consistency and control sufficient to power bulk production.

Where to Access the Models

Both models have begun rolling out:

The fastest way for developers to evaluate the technology is via Speech Generation in Google AI Studio. Google has designed the interface as a voice design studio: users can craft voices from scratch via prompts or clone their own voice, then load them directly into a two-speaker script editor for line-by-line directing.

Safety: Consent Verification and Invisible Watermarking

The primary danger of voice cloning is impersonation. Google applies a three-layer defense:

• Consent verification: Prior to voice replication, the voice owner must verbally record a consent statement that matches the reference audio sample.
• SynthID watermarking: All audio synthesized by Gemini Audio models embeds Google's SynthID watermark. Inaudible to human ears, it is woven directly into the audio structure and can verify AI generation downstream to counter misinformation.
• C2PA credentials: Replicated voices include C2PA content credentials, an open industry provenance standard documenting which entity and tools generated the content.

Additional safety specifications can be found in the Gemini 3.8 Audio Model Card.