tts-1

TTS-1 is OpenAI's text-to-speech model optimized for speed. It receives the words to be spoken and produces audio, rather than a transcription or a written answer. The manufacturer's examples include narrating written articles, producing speech in several languages and playing audio as it arrives. TTS-1 is the lower-latency choice compared with TTS-1 HD; the HD model emphasizes output quality. The useful decision is therefore whether prompt playback or the higher-quality narration profile matters more for the task.

The speech-generation workflow is built around the model, the input text and the selected voice. OpenAI's reference limits the text input to 4096 characters, a character limit rather than a token context window. Prepare the actual script before synthesis: write the sentences the listener should hear, resolve abbreviations and decide how names, quantities and headings should be spoken. For a long article, organize sections into independently understandable passages instead of treating a long document as a single speech request.

OpenAI's current model-specific voice list for TTS-1 is alloy, ash, coral, echo, fable, onyx, nova, sage and shimmer. These are preset voices; they are not the full voice set of every newer speech model. The voices are optimized for English, while multilingual speech is supported by providing the input in the language to be spoken. Choose a voice using a representative passage with the same vocabulary and pacing as the final script. Keep that choice consistent across related clips when continuity matters.

Audio can be streamed in chunks so playback begins before the entire file has finished generating. This is streaming of speech created from text, not a two-way spoken conversation. The documented output formats are MP3, Opus, AAC, FLAC, WAV and PCM, with MP3 as the default. MP3 and AAC provide compressed audio; Opus is suited to streaming and communication; FLAC is lossless. WAV avoids compressed-audio decoding, and raw PCM contains 24 kHz, 16-bit signed little-endian samples without a file header. OpenAI recommends WAV or PCM when response time is the priority.

The speech reference provides a speed setting from 0.25 to 4.0, with 1.0 as the default. The additional natural-language instructions field does not work with TTS-1, and event-style SSE output is not supported for this model. Its documented controls are the script, voice, speed and audio format; newer models' controls for accent, emotion or whispering should not be attached to this edition. When preparing a clip, choose the speed to serve intelligibility and then inspect the rendered words and pauses rather than assume the script's character count fixes an exact duration.

The language guide includes English, Russian, Ukrainian, French, German, Spanish, Portuguese, Italian, Arabic, Chinese, Japanese, Korean, Hindi, Kazakh and many other languages. Speech follows the language of the supplied text. A multilingual narration brief therefore starts with the translated script and the intended audience; text-to-speech is a rendering step, not an instruction to translate or rewrite it. For a practical announcement, prepare the exact message and names. For educational narration, make section boundaries and definitions clear in the words that will be read.

A finished artifact can be a short spoken instruction, a narrated passage or an audio segment for a larger edit. Review pronunciation, completeness, pacing and the chosen format using the actual output. Preserve the script together with the audio so later corrections can be localized to the affected passage. OpenAI's guidance requires listeners to be clearly told that the voice is AI-generated. That disclosure belongs to the listening experience; it does not turn TTS-1 into a human-voice recording or add voice-cloning controls to the model.