OpenAI TTS

OpenAI's Text-to-Speech API is a developer-facing speech-synthesis offering delivered through the same `/v1/audio/speech` endpoint and SDK ecosystem as the rest of the OpenAI platform. It spans three model generations: the original tts-1 (latency-optimized) and tts-1-hd (quality-optimized), both priced per character, and the March-2025 gpt-4o-mini-tts, a token-priced, "steerable" model that lets developers prompt not just what is said but how, controlling accent, emotion, tone, pacing and whispering via a free-text instructions field. It supports 13 built-in voices (with marin and cedar billed as the highest quality), outputs MP3/Opus/AAC/FLAC/WAV/PCM, streams audio via chunked transfer, and follows Whisper's broad multilingual coverage (90+ languages). The headline pitch is integration speed: if you are already on OpenAI APIs, TTS is a drop-in REST call with predictable pricing.

The product's center of gravity is convenience rather than best-in-class voice fidelity. Independent and community comparisons consistently frame OpenAI TTS as "good enough" voice quality that ships fast, versus ElevenLabs' superior emotional range and voice cloning. OpenAI reports its newest snapshot delivers roughly 35% lower word error rate on Common Voice and FLEURS, and one third-party preference test put OpenAI TTS on top at a 42.93% preference rate with 87.13% pronunciation accuracy. Where it loses is real-time latency and voice-cloning depth: benchmarks show tts-1-hd P50 latency over one second (not viable for live voice agents), and time-to-first-audio that trails ElevenLabs and Cartesia. Custom/cloned voices exist but are gated to eligible organizations with a consent-recording requirement.

For teams building voice agents at the cutting edge of latency, OpenAI now steers them toward the Realtime API and its newer realtime TTS models rather than the classic speech endpoint. The classic TTS API is best understood as a pragmatic batch/near-real-time synthesis tool: simple, cheap, well-documented, multilingual, and tightly integrated with the broader OpenAI stack, at the cost of the per-request 4,096-character limit (a recurring forum complaint), no fine-grained SLA on the Standard tier, and voices that sound competent but less expressive than dedicated voice specialists. Reliability and SLAs (99.9% uptime) are only contractually guaranteed on Scale/Priority tiers, not the default Standard tier.