Voice AI Stack Explained: Models, Frameworks & Platforms (2026)

The Voice AI stack is categorized by the specific roles that different products play, ranging from individual models to fully finished agents.

The Layers of the Stack

• Models (Ingredients): These are sold to developers via APIs. They include speech-to-text (converting voice to words), language models (deciding what to say), and text-to-speech (converting text back to voice). Alternatively, speech-to-speech models combine these steps to hear and speak directly. Small ingredients like voice activity detection, turn detection, and noise filtering determine when an agent listens and speaks.
• Plumbing: This layer manages how audio travels in and out. It consists of WebRTC for apps and browsers, and telephony for phone calls. This layer directly impacts audio quality and delay.
• Open-Source Frameworks: These allow developers to assemble the pipeline themselves using free code. Developers have full control to swap components and run them on their own servers, but they must invest engineering time and manage any bugs.
• Managed Platforms: These are rented builders where agents can be set up quickly via a dashboard or API. Users trade some control for speed and typically pay per minute. When evaluating costs, it is important to compare the all-in cost per minute, as some platforms bill underlying models and phone lines separately while others bundle them.
• Verticals: These are finished agents built for specific industries, requiring no building from the buyer.
• Evals (Evaluation Tools): These tools test and monitor the performance of built agents.
• Hardware: Physical devices that have voice capabilities built directly into them.

Shifting Borders and Choosing a Starting Point

The boundaries between these layers are fluid, and single companies often play multiple roles. For example, companies that started in infrastructure or single-model APIs have expanded to offer open-source frameworks, speech-to-text, and full agent platforms.

Deciding where to start depends on a team's specific goals:

• Teams with engineers who want to own the pipeline should look at frameworks.
• Those needing a working agent quickly should look at platforms.
• Those wanting a completed job for a specific industry should look at verticals.