Google releases Gemini 3.8 speech synthesis models

Summary

Google released two new Gemini 3.8 text-to-speech models (Flash and Flash-Lite), featuring over 2,000 voices, custom voice cloning from a 30-second sample, and multi-speaker dialogue synthesis. Simon Willison showcased these models through a browser-based application arena he developed with GPT-6 Astra.

Why it matters

Near-instant generation (1 minute 18 seconds of audio in about 20 seconds), extremely low cost (2.74 cents), and multi-character dialogue support lower barriers for interactive storytelling, accessibility tools, and prototyping. An open CORS policy and a bring-your-own-key model provide developers with immediate access without backend requirements, accelerating experimental work. Custom voice cloning from just a 30-second audio recording raises expectations in the text-to-speech field, both creatively and ethically.

Highlights

• A library of over 2,000 voices and custom voice cloning from a 30-second sample
• A multi-speaker dialogue API with per-character voice and style instructions
• Approximately 20-second generation time and a cost of 2.74 cents for 1 minute 18 seconds of Flash TTS audio

Critical take

The article is essentially a product demo dressed up as a blog post. While the 'vibe-coded' application arena and pelican sketch are charming, it says little about real-world latency, voice fidelity, or the practical risks of voice cloning from a 30-second recording. Without independent quality benchmarks or comparisons to competing text-to-speech solutions, it is difficult to contextualize the 2.74-cent price point without knowing the upper limit of audio quality.