The New Gemini Text to Speech

Gemini Text-to-Speech: A Deep Dive

Key Concepts:

• Gemini Text-to-Speech (TTS): Speech generation models built on the Gemini series LLMs, offering high-quality speech synthesis with nuanced control.
• Gemini 2.5 Pro/Flash: The specific Gemini models currently powering the TTS functionality (in preview). Flash is noted for speed and adherence to instructions.
• Audio Modality: The configuration setting within the Gemini API that specifies audio generation.
• Director Nodes: Specific prompting techniques to control style, accent, and pace in generated speech.
• Audio Profile: Defining the character's core identity and archetype within a prompt.
• Scene Description: Establishing the environment and emotional vibe within a prompt.
• Agentic Clothing Platform (Blink): A platform discussed as a sponsor, enabling rapid application development from text prompts.

1. Introduction & Capabilities

The video focuses on Gemini Text-to-Speech (TTS), a new speech generation capability built on the Gemini series of large language models (LLMs). This system aims to rival the quality of dedicated speech models like those from 11 Labs, but with the added benefit of being driven by an LLM. This allows users to describe desired speech effects – dramatic pauses, specific emotions, accents – directly in the prompt, and the model attempts to generate speech accordingly. The presenter highlights the potential for this technology to “transform how you create audio content.” A key advantage is the elimination of the need for specialized tokens for special effects, and the ability to generate multiple speakers easily.

2. Practical Implementation & API Access

The video provides a practical guide to using the Gemini 2.5 TTS models. It’s important to note these are currently in preview and are built on the previous generation Gemini, not the latest Gemini 3. To access the functionality, a Gemini API key is required (a paid key is recommended to avoid rate limits). The Google Generative AI SDK (version 1.16 or higher) is also necessary. The presenter provides a link to a notebook in the video description for easy implementation.

The basic process involves:

• Importing the necessary SDK.
• Providing the API key and creating a generative AI client.
• Specifying the model ID (either Gemini 2.5 Flash Preview TTS or Gemini 2.5 Pro Preview TTS).
• Setting the configuration to use the audio modality.
• Providing a text string containing both instructions and the speech content.

3. Gemini 2.5 Flash vs. Pro

The presenter’s experiments suggest that the Gemini 2.5 Flash model performs better than the Pro model for TTS. Flash is significantly faster and appears to adhere to instructions more closely. The recommendation is to experiment with both to determine which best suits specific needs.

4. Prompting & Control

A core strength of Gemini TTS is its responsiveness to natural language prompting. Users can directly describe the desired style, tone, accent, and pace. The system also supports specific instructions within the text string, such as pauses ("wait for 5 seconds"), which the model accurately implements. The system supports a wide range of languages (over 24, including Arabic and Hindi).

The presenter introduces a structured prompting approach:

• Audio Profile: Defines the character's core identity.
• Scene Description: Sets the environment and emotional tone.
• Director Nodes: Provide precise guidance on style, accent, and pace.

This approach allows for highly crafted speech outputs. An example is given of a scene with "massive vibes" and instructions to create an energetic, fast-paced delivery.

5. Multi-Speaker Capabilities

Gemini TTS allows for the generation of speech with multiple speakers. This can be achieved by defining different speakers and assigning them specific voices from a pre-built list. The example provided demonstrates a conversation between two speakers with contrasting personalities (tired/bored vs. excited/happy). The functionality is described as being similar to that of NotebookLM.

6. Limitations & Areas for Improvement

While impressive, the system isn’t perfect. The presenter acknowledges that Gemini TTS, like other LLMs, struggles with humor. An example of a comedy routine generated by the model demonstrates that while not bad, it lacks genuine comedic timing and delivery.

7. Voice Options & Language Support

The system currently offers around 30 different voice options. It supports up to 24 languages for speech output, including many European, Eastern, and Asian languages, a significant advantage over systems that often neglect these languages.

8. Technical Specifications & Context Window

The model has a context window of 32,000 tokens, which is less than the 1 million token context window of the base Gemini model.

9. Pricing

The pricing structure is as follows:

• $0.50 per 1,000 input tokens (text).
• $10 per 1 million output tokens (audio).
• The Pro version is double the price.
• Batch processing reduces the price to half.

10. Sponsor Segment: Blink

The video includes a sponsored segment featuring Blink, an agentic clothing platform that allows users to create full-stack applications from simple text prompts, including features like authentication and payment processing. Blink utilizes models like Nano Banano for image generation and offers native integrations for rapid development.

11. Conclusion & Future Outlook

The presenter concludes that Gemini TTS is a promising technology with the potential to revolutionize AI voice and speech applications. They encourage viewers to experiment with the system and share their experiences. The presenter predicts that speech and voice will be a major area of focus for frontier AI labs in 2026.

Notable Quote:

“In a world where artificial intelligence has changed everything, one API will transform how you create audio content. Gemini text to speech coming to a developer near you this summer.” – Opening narration.

Data & Statistics:

• Over 24 languages supported.
• 30 different voice options available.
• Context window of 32,000 tokens.
• Pricing: $0.50/1,000 input tokens, $10/1 million output tokens (Flash); double for Pro.