The TTS Landscape

A startup founder in Austin had a problem. Her app helped elderly patients manage medications, but many of her users could not read the small text on their phone screens. She needed the app to speak -- clearly, warmly, in a voice that did not sound like a GPS navigator from 2008. She evaluated eleven TTS providers over two weeks, burned through trial credits, and ultimately chose OpenAI's API because a single parameter change let her switch from a professional narrator voice to a warm conversational one. The entire integration took four hours.

That is the state of text-to-speech in 2025. The technology has crossed the uncanny valley. The voices sound human. The real engineering challenge is no longer "can the machine talk" but "which voice, at what quality, for what cost, and how fast."

What You Will Learn

1. The three OpenAI TTS models and when to use each
2. All available voices and their characteristics
3. Quality, latency, and cost tradeoffs that affect production decisions
4. How TTS fits into larger application architectures

Three Models, Three Tradeoffs

OpenAI offers three text-to-speech models, each designed for a different point on the quality-speed-cost spectrum.

gpt-4o-mini-tts

This is the newest and most capable model. It supports 13 voices and -- critically -- accepts an instructions parameter that lets you control accent, tone, emotion, and speaking style with natural language. Tell it to "speak with a warm Southern accent" or "sound excited about the product launch" and the output shifts accordingly.

The 13 voices are: alloy, ash, ballad, coral, echo, fable, nova, onyx, sage, shimmer, verse, marin, and cedar. Each has a distinct vocal character -- some deeper, some brighter, some more conversational, some more authoritative.

tts-1

The standard model. Supports 6 voices (alloy, echo, fable, onyx, nova, shimmer) plus 3 additional ones (ash, coral, sage) -- 9 total. No instructions parameter. Lower latency than tts-1-hd, which makes it suitable for real-time applications where speed matters more than maximum audio fidelity. At high listening volumes or through studio headphones, you may notice slight artifacts compared to the HD variant, but for most consumer applications -- phone speakers, earbuds, car audio -- the difference is negligible.

tts-1-hd

Same 9 voices as tts-1, but rendered at higher quality. The audio is smoother, with less compression artifacts and more natural intonation at the edges of phrases. The tradeoff is higher latency and higher cost. Use this when audio quality is the primary concern -- audiobook narration, professional voiceovers, accessibility features where clarity is non-negotiable.

Voice Selection Guide

Choosing the right voice is a product decision, not a technical one. The wrong voice creates cognitive dissonance -- a cheerful voice delivering bad medical news, a monotone voice trying to teach children.

Quality vs Latency vs Cost

Every production decision involves a triangle of constraints.

Latency matters when users are waiting for a response in real time -- voice assistants, interactive tutorials, accessibility features that read UI elements aloud.

Quality matters when the audio is the product -- audiobooks, podcast intros, professional narration, branded voice experiences.

Cost matters when you are generating high volumes -- reading every email aloud for an accessibility app, generating audio for thousands of articles, running a service with millions of daily active users.

For most applications, start with gpt-4o-mini-tts and the instructions parameter. The expressiveness and control it provides outweigh the marginal cost difference for all but the highest-volume use cases. Drop to tts-1 only if latency benchmarks in your specific deployment show a meaningful difference.