What for: cheap, fast voiceover through the OpenAI API (gpt-4o-mini-tts), with delivery steered by a plain-text instruction.

What it is and who it's for

OpenAI TTS is speech synthesis through the OpenAI API (the gpt-4o-mini-tts model and others): cheap, fast, with intonation controlled by a plain instruction. For a developer who needs voiceover "in the flow" of an app, inside the same ecosystem as the rest of their OpenAI calls. What sets it apart: the reading style is set by a text instruction (instructions) rather than by presets.

Key features

Several built-in voices; the instructions parameter controls tone and emotion.
Low price and fast synthesis.
Audio streaming so playback starts quickly.
Output formats mp3/opus/aac/pcm and others.
One ecosystem with the rest of the OpenAI API (one key, one SDK).
Support for many languages, including Russian.

Get started in 5 minutes

Grab an OpenAI key (platform.openai.com) and top up your balance.
Synthesize speech with a single request:

curl https://api.openai.com/v1/audio/speech \
  -H "Authorization: Bearer $OPENAI_API_KEY" -H "Content-Type: application/json" \
  -d '{"model":"gpt-4o-mini-tts","voice":"alloy","input":"Привет, мир","instructions":"говори бодро"}' --output out.mp3

Change voice and instructions until you land on the delivery you want.

When to take it, and when not

✅ Take it if: you already use OpenAI and need cheap, fast voiceover of acceptable quality inside one SDK.
✅ Take it if: you care about steering the delivery with a text instruction.
❌ Skip it → "ElevenLabs" is better if you need the most natural sound, voice cloning or dubbing.
❌ Skip it → "Kokoro TTS" is better if you need privacy/offline and zero cost per minute.

The honest price

Paid per character/token — but it is one of the cheapest quality TTS options on the market; there is no free tier. Exact amounts are on the provider's side and change over time.

Gotchas

Fewer voices, and less "characterful" ones, than ElevenLabs; no cloning.
Intonation control via instructions works, but is not perfectly predictable.
Data goes to OpenAI's cloud — not suitable for private scenarios.
In Russian there are occasional slips with stress and numbers — check and fix the input text.

🤖 Prompt accelerator

"I'm doing Russian-language voiceover with OpenAI TTS (gpt-4o-mini-tts). Help me pick a voice for , write instructions for the delivery I want (pace, emotion), prepare the text so numbers and abbreviations read correctly, and give me an example call with streaming in <Python/Node> that saves to a file."