GPT-4o-mini-TTS is OpenAI's text-to-speech model for applications that already have written text and need controllable spoken audio. It accepts text only and returns audio only. Unlike a general chat model, it is not designed to analyze images, transcribe incoming speech, maintain a long conversation history, or return a text answer.
The model's main advantage over older preset-style TTS workflows is its instructions field. Instead of selecting only a voice and speed, developers can describe the intended delivery in ordinary language: calm or excited, conversational or formal, softly spoken or energetic, with guidance for accent, pacing, intonation, and emphasis. This makes the OpenAI GPT-4o-mini-TTS API useful for narration, product onboarding, accessibility audio, game dialogue, and spoken responses generated from an application's text output.
| Specification | GPT-4o-mini-TTS |
| Model string | gpt-4o-mini-tts |
| Input modality | Text |
| Output modality | Audio |
| Maximum input | 2,000 input tokens per request |
| Built-in voices | 13 |
| Output formats | MP3, Opus, AAC, FLAC, WAV, PCM |
| Delivery | Complete audio response or streamed audio |
| Adjustable speed | 0.25 to 4.0; default 1.0 |
| GPTProto pricing | $0.42 text input / $8.40 audio output per 1M tokens |
Choose the Right Voice and Audio Format
GPT-4o-mini-TTS currently documents the built-in voices alloy, ash, ballad, coral, echo, fable, nova, onyx, sage, shimmer, verse, marin, and cedar. OpenAI recommends trying marin or cedar first for quality-focused output. The voices can generate speech in multiple languages, but they are optimized for English, so test names, acronyms, numbers, and local pronunciation with real scripts before release.
Voice and format solve different problems. The voice determines the base sound; the instruction controls delivery; the response format determines how the audio fits the product.
| Format | Best fit |
| MP3 | General web playback, downloads, and compact files |
| Opus | Internet streaming and communication applications |
| AAC | Mobile apps and video publishing workflows |
| FLAC | Lossless storage, editing, and archiving |
| WAV | Low-latency playback and audio editing without compressed-audio decoding |
| PCM | Raw 24 kHz, 16-bit little-endian audio for custom playback pipelines |
Use MP3 as a practical default. Choose Opus when bandwidth matters, FLAC when lossless storage matters, and WAV or PCM when avoiding decoding overhead is more important than file size.
How to Write a GPT-4o-mini-TTS Prompt
A useful GPT-4o-mini-TTS prompt describes four things: the speaker's role, the listener, the emotional intent, and the delivery. Keep the spoken script in the input field and put voice direction in the instructions field. This separation makes prompts easier to reuse and test.
| Application | Example instructions prompt |
| Customer support | Speak calmly and reassuringly. Use a warm conversational tone, moderate pace, and gentle emphasis on the next step. Do not sound overly cheerful. |
| Product onboarding | Sound friendly, clear, and confident. Keep the pace brisk, pause briefly after each action, and emphasize button names without sounding like an advertisement. |
| Short-form narration | Use an intimate documentary style with controlled energy. Begin softly, build curiosity through the middle, and slow down for the final sentence. |
| Accessibility reading | Read clearly at a measured pace. Pronounce every number and abbreviation distinctly, avoid dramatic emotion, and pause between headings and body text. |
Change one variable at a time when testing a GPT-4o-mini-TTS API voice. First select the base voice, then refine tone and pacing, and finally test pronunciation. This makes it easier to identify whether a problem comes from the voice choice, the instruction, or the source text.
Common GPT-4o-mini-TTS API Applications
-
Narration and voiceovers: Convert articles, product videos, lessons, and short scripts into consistent spoken audio.
-
App guidance: Read onboarding steps, alerts, navigation cues, or generated summaries aloud.
-
Voice-agent output: Turn a text response from an assistant into streamed speech. Use a Realtime API workflow instead when the product needs full two-way, speech-to-speech interaction.
-
Accessibility: Add spoken versions of messages, documents, and interface content for users who prefer or require audio.
-
Multilingual delivery: Generate speech from supported-language text, while testing each target language because the built-in voices are optimized for English.
Plan Long Scripts Before Sending Requests
The maximum input is 2,000 tokens per request, not a 128K context window. Split longer material at sentence or paragraph boundaries. Reuse the same voice, instruction, speed, and response format for every segment, and avoid cutting a sentence, number, or proper name across two requests. After generation, listen to the joins for repeated pauses, inconsistent emphasis, or pronunciation changes.
For interactive playback, request streaming so the application can start playing audio before generation finishes. For prerecorded narration, generate complete files and run a quality check before publishing. In either case, clearly disclose to end users that the voice is AI-generated rather than human.
