The New Gemini Text to Speech

Quick Overview

Google AI Studio has released significant enhancements to its Gemini 2.5 Flash and Gemini 2.5 Pro Text-to-Speech (TTS) preview models, introducing improved expressivity, precision pacing, and seamless dialogue capabilities, with the Flash model offering faster generation than the Pro version.

Key Points: The key improvements in Gemini 2.5 TTS models include enhanced expressivity (richer tone versatility), precision pacing (smarter context-aware speed adjustments), and seamless dialogue (consistent character voices). The Gemini 2.5 Flash TTS preview model is optimized for performance and low latency, while the Gemini 2.5 Pro TTS preview model is optimized for higher quality outputs. The Gemini TTS models can automatically detect input language and support 24 languages, including European and Asian languages. Users can control style, tone, accent, and pace using natural language prompts, demonstrated by generating a spooky whisper and a fast-paced podcast debate. Multi-speaker audio generation is supported by specifying different objects for each speaker, allowing assignment of distinct voices like 'Puck' (enthusiastic) and 'Fenrir' (skeptical). The paid tier pricing for Gemini 2.5 Flash TTS is $0.50 per 1M text tokens (input) and $10.00 per 1M audio tokens (output), while the Pro version is $1.00 (text) and $20.00 (audio) for batch processing. The TTS models have a context window limit of 32k tokens, which is smaller than the 1 million token context window of the base Gemini models.

Context: This video details the significant enhancements made to Google's Gemini 2.5 Text-to-Speech (TTS) preview models, specifically Gemini 2.5 Flash and Gemini 2.5 Pro. The presenter walks through a Google Colab notebook demonstrating how to use the updated API to generate more expressive, context-aware, and multi-speaker audio content, highlighting new prompting capabilities and pricing structures.

Detailed Analysis

The video showcases the new features of the Gemini 2.5 Text-to-Speech (TTS) preview models, which are built upon the larger Gemini language model. Key improvements focus on better control over audio output: Enhanced expressivity allows for richer tone versatility and stricter adherence to style prompts; Precision pacing enables smarter, context-aware speed adjustments; and Seamless dialogue ensures consistent character voices in multi-speaker scenarios. The presenter demonstrates these capabilities using Python code in a Google Colab notebook. First, a basic audio generation is shown, requiring the configuration parameter to request audio output. The models support 24 languages, automatically detecting the input language using BCP-47 codes. Advanced prompting allows users to dictate style, tone, accent, and pace, as demonstrated with a 'spooky whisper' and a fast-paced comedy show script with contrasting speaker personalities (enthusiastic vs. skeptical). The multi-speaker feature requires configuring with specific objects, assigning prebuilt voices like 'Puck' and 'Fenrir' to different speakers. The presenter notes that the Flash model (optimized for performance) is faster in experiments than the Pro model (optimized for quality). Finally, pricing tiers are reviewed: Gemini 2.5 Flash standard tier charges $0.50/1M text tokens and $10.00/1M audio tokens in the paid tier, while the Pro version is $1.00/1M text and $20.00/1M audio (batch pricing). A limitation noted is that TTS sessions have a smaller context window limit of 32k tokens compared to the 1 million tokens supported by the base Gemini models.