Mountain View, California – Google has expanded its Gemini Audio suite with the introduction of Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS. These next-generation text-to-speech (TTS) models are presented as the company's most expressive audio generation capabilities to date, providing developers with enhanced tools for creating dynamic and nuanced vocal performances.
Two Tiers for Diverse Production Needs
The latest release segments its TTS offerings into two distinct models, each tailored for specific use cases while sharing advanced direction controls:
- Gemini 3.8 Flash TTS: This model is engineered for deep creative control and character development. It targets applications such as gaming, immersive audiobooks, podcasts, and interactive media. Flash TTS enables precise command over acting cues, pacing, subtle dialect shifts, and responsive backchanneling.
- Gemini 3.8 Flash-Lite TTS: Optimized for high-volume, cost-efficient scaling, Flash-Lite TTS is designed for tasks like dubbing, extensive audio content creation, and expressive voice agents. It offers fine-grained adjustments for tone, delivery speed, and overall expressive nuance.
Both models are now accessible via the Gemini API and Google AI Studio, with enterprise-level API access through Gemini Enterprise slated for future availability. Currently, the models are exclusively available through API, without options for self-hosting via open weights.
Revolutionizing Voice Creation from Text Prompts
Moving beyond the 30 original voices offered by previous Gemini TTS iterations, the 3.8 release introduces a vastly expanded and more flexible voice system:
- Generative Voice Design: Flash TTS facilitates the creation of entirely new voices based on text prompts that describe specific roles, accents, and vocal characteristics. This functionality supports over 100 languages and dialects, demonstrated with diverse examples such as a Melbourne DJ, a robotic monotone, and a Japanese dragon.
- Extensive Voice Library: Developers gain access to a comprehensive library of more than 2,000 production-ready voices, encompassing a wide range of regional variations, including Mexican Spanish, Quebec French, and Scots English.
- Custom Voice Management: User-generated custom voices can be saved and consistently reused across multiple projects, ensuring minimal vocal drift.
- Upcoming Voice Remixing: A future enhancement will allow users to modify a library voice's timbre, pitch, pace, and accent directly through prompts.
Granular Performance Direction and Dialogue Management
Both models permit the embedding of stage directions directly within the script. Gemini can also intelligently guide delivery based on natural cues present in the text, ensuring a fluid and contextually appropriate performance:
- Long-Form Consistency: The models maintain consistent voice quality, pacing, and timbre across audio generations lasting several hours.
- Native Two-Speaker Staging: A single script can drive multi-turn conversations featuring distinct, separated voices for each participant.
- Vocal Bursts: Non-verbal expressions, such as
<laughs>,<sigh>, and<gasp>, can be integrated to add conversational texture and realism. - Backchanneling: Active-listening interjections like
|mhm|and|yeah|provide control over reaction timing and comedic delivery.
Safety Measures and Ethical Controls
Google has incorporated robust safety features, particularly concerning voice replication:
- Voice Replication Safeguards: Creating a consistent vocal profile from a 30-second audio sample requires explicit verbal consent from the voice owner, which is matched against the reference speaker. The sampled voice must either belong to the user or be one for which they possess usage rights.
- Audio Watermarking: Every audio clip generated by Gemini Audio models includes an imperceptible SynthID watermark embedded directly into the output.
- Content Credentials: Replicated voices are also equipped with C2PA content credentials, further enhancing transparency and traceability. For a comprehensive overview of its safety methodology, Google directs users to the Gemini 3.8 Audio model card.
Industry-Leading Performance Benchmarks
The new models have demonstrated strong performance in independent evaluations:
- Hume AI Voice Design Benchmark: Flash TTS achieved the top position overall with a score of 71.4, according to Hume AI.
- Accent Modeling: Flash TTS led in accent modeling, scoring 60.8.
- Hume AI Overall Quality Index: Flash TTS ranked first, with Flash-Lite TTS securing the second position.
- Voice Arena Blind Preference: Both models garnered top preferences in blind tests across multiple languages, including Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish, and Hindi.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
