# Why voice agents are about to be everywhere: Dylan Fox, CEO of AssemblyAI

## Executive summary

AssemblyAI, through its CEO Dylan Fox, discusses the rapid adoption of conversational voice agents and the critical role of high-quality Text-to-Speech (TTS) models. The company launched TTS-1, a proprietary model built for conversational use cases. AssemblyAI emphasizes its strategy of owning the entire 'voice loop' (Speech-to-Text and Text-to-Speech) and leveraging a unified inference infrastructure. This approach allows developers to build scalable, low-cost voice applications with uncapped rate limits, addressing the core need for natural, intelligent voice interactions in both B2B and consumer environments.

## Key takeaways

- The Voice as a Tell: Users judge the intelligence and utility of a voice agent based on the quality of the voice. If the voice is poor, users are likely to hang up, regardless of the underlying AI capability. (0:00, 9:28)
- TTS-1 for Conversational Agents: AssemblyAI launched TTS-1, a text-to-speech model specifically trained on proprietary conversational data to ensure a natural, conversational sound, moving beyond the quality of traditional audiobooks or podcasts. (2:36)
- Infrastructure Advantage of the Full Voice Loop: By serving TTS-1 on the same proprietary inference stack used for Speech-to-Text (STT), AssemblyAI offers a significant advantage: a frontier-quality model at a low cost ($15 per 1 million characters) with uncapped rate limits and high reliability. (12:56)

## Technical details

- TTS-1 Model Specifications: TTS-1 is AssemblyAI's first TTS model, purpose-built for conversational voice agents. It is priced at $15 per 1 million characters and is claimed to be the lowest-cost frontier TTS model for voice agents, optimizing quality, cost, and latency. (2:36)
- Voice Infrastructure Scale: The platform handles over 2 billion API calls monthly and over 100 million conversations weekly. This scale necessitated building proprietary inference software and a runtime, which now powers both STT and TTS services. (12:56)
- Future Voice Capabilities: AssemblyAI is developing a live speech-to-speech duplex model, scheduled for release in early December, which will run on the same unified inference stack. (12:56)

## Practical implications

- For developers building voice agents, the ability to use a unified inference stack means they do not need to worry about rate limits or compute capacity when scaling their applications.
- The focus on co-designing models based on specific use cases (e.g., quick transactional B2B vs. long consumer conversations) suggests that developers must tailor their voice agent architecture to the intended user workflow and industry risk profile.
- The availability of instant voice cloning (with consent) allows developers to deeply integrate the AI voice into a brand's specific identity, enhancing user experience.

## Topics

Voice AI, Text-to-Speech (TTS), Speech-to-Text (STT), Conversational AI, Inference Infrastructure, API Development, AssemblyAI Website

Source: https://www.youtube.com/watch?v=ulJVhhkM-ao