SpaceXAI: Grok Voice TTS 1.0

Grok Text to Speech is the Grok voice family’s text-to-audio capability. The manufacturer documents expressive built-in speakers, inline vocal cues and output formats suited to media playback or telephone audio. Its input is the transcript to be spoken, together with voice and language choices. The resulting recording is speech from that transcript; speech recognition and conversational speech-to-speech models have separate tasks.

Voice, language and spoken text

The producer catalogue gives each voice a delivery profile. Ara is warm and friendly, Eve is upbeat with a British accent, and Rex is described as clear and confident. Voice and language are independent in the documented family: built-in speakers can use each supported language choice. The guide lists twenty language-code options, including regional variants, and also supports automatic language detection. Alongside built-in speakers, the producer documents reusable custom voices created from a reference recording and selected by voice_id for complete or streaming synthesis. Reference audio may be up to 120 seconds; 90–120 seconds of clean, single-speaker speech is recommended to capture delivery and intonation. Keep the reference, created voice profile and generated recording as separate artifacts, and match the reference performance to the intended script.

Prepare the final spoken wording, preserving the intended language of names and quoted passages. For a narrated guide, a useful line is: “Choose a topic, listen to the short introduction, and open the section you want to explore.” If a brand name or specialist term needs a particular reading, the manufacturer’s replacement controls can supply a respelling or IPA pronunciation without changing the displayed source text. A separate text-normalisation option converts written numbers, abbreviations and symbols towards their spoken form.

Performance cues

Inline tags place an event at a particular point, such as a pause, laugh, sigh or breath. Wrapping tags apply delivery to a complete phrase, for example whispering, emphasis or a slower passage. Put the cue where the meaning calls for it and keep the words readable. The guide recommends complete-phrase style spans and punctuation rather than a stack of unrelated tags.

For an expressive character line, prepare the spoken sentence and the performance shift together. For an explanation, retain a clear rhythm that helps the listener follow the steps. The documented speech-rate range is 0.7 to 1.5, with 1.0 as normal delivery. A speed choice changes the spoken pace; it does not change the words the application should communicate.

Audio and timings

The producer documents MP3, WAV, raw PCM, mu-law and A-law output. Supported sample-rate choices span 8 kHz to 48 kHz, with specific listed rates between those values. Its default MP3 configuration is 24 kHz at 128 kbit/s. Use an audio representation that the intended player, editing stage or telephone path can interpret.

Optional character timings associate the returned character sequence with start and end times. Without pronunciation replacements, that sequence mirrors the input; replacements can change it to the spoken text. They can support a read-along display, highlighting or alignment with a visual performance. In the complete-response workflow, requesting timings changes the output into an envelope containing encoded audio and timing data. Treat the timings and audio as a paired result, especially when text normalisation expands a written symbol into several spoken words.

Complete or streaming generation

The manufacturer offers a complete-request path and bidirectional streaming, in which text fragments arrive and audio chunks return as the utterance is generated. It documents a 60,000-character limit for a complete request and for an individual streamed text fragment. Completing or clearing an utterance has an explicit role in the streaming workflow.

For longer narration, choose coherent passage boundaries and retain the script with each result. For interactive playback, distinguish text sent for synthesis from audio already heard by the listener. The useful deliverable is a voiced script with intentional performance cues, an appropriate audio format and timing data when the application needs it. Keep the script, selected voice, performance cues and output settings with the audio so later revisions preserve the intended delivery.