Do Speech-to-Speech Models Work in Production? (Full Transcript)

A survey asks voice AI teams whether speech-to-speech models handle production workflows, topic shifts, and tool calling reliably.

Download Transcript (DOCX)

Have you shipped speech-to-speech to production

Added on 11/05/2026

Speakers

Add new speaker

[00:00:00] Speaker 1: In the world of voice models, there are two kinds which are very popular. One is you cascade speech to text model, then an LLM and then text to speech model. And the number two way would be a dedicated speech to speech model. Now the big question is, have you tried speech to speech model? Have you seen if they actually work in production? When you switch topics, when you do tool calling, do they actually do the job that they are supposed to? And if you did try, did it actually make to production? So that's why today we are surveying all the voice builders, founders, engineers and the first 200 people get $25 of gift card and early results to the survey. Sign up below.

Summary

The speaker contrasts two popular voice AI architectures: a cascaded pipeline of speech-to-text, an LLM, and text-to-speech, versus dedicated speech-to-speech models. They invite voice builders, founders, and engineers to share whether speech-to-speech systems work reliably in production, particularly for topic changes and tool calling. The first 200 survey respondents will receive a $25 gift card and early access to survey results.

Copy

Download

Title

Surveying Production Readiness of Speech-to-Speech Voice Models

Copy

Download

Keywords

voice models Remove

Remove

speech-to-speech Remove

Remove

speech-to-text Remove

Remove

text-to-speech Remove

Remove

LLM

Remove

voice AI

Remove

production readiness Remove

Remove

tool calling Remove

Remove

survey

Remove

voice builders Remove

Remove

Copy

Download

Key Takeaways

Two common voice AI approaches are cascaded STT→LLM→TTS pipelines and dedicated speech-to-speech models.
The survey focuses on whether speech-to-speech models are reliable in real production settings.
Key concerns include performance during topic switches and tool-calling workflows.
Voice builders, founders, and engineers are invited to participate.
The first 200 respondents receive a $25 gift card and early survey results.

Copy

Download

Sentiments

Neutral: Informative and promotional tone, posing practical questions about model reliability and offering an incentive for survey participation.

Copy

Download

Enter your query

{{ secondsToHumanTime(time) }}

Back

Forward

{{ Math.round(speed * 100) / 100 }}x

{{ secondsToHumanTime(duration) }}

Select Audio file