xAI launches Grok STT and TTS voice AI APIs
xAI launches `Grok Speech to Text` and `Grok Text to Speech`, two APIs for transcribing audio and generating synthetic voices from any application. They include speaker identification, timestamps, more than 25 languages and expression controls, with pricing based on audio hours and character count.
xAI has launched two independent APIs for adding voice to any application: Grok Speech to Text converts audio into text, while Grok Text to Speech turns text into a synthetic voice. The company says both use the same technology that powers Grok Voice, Tesla vehicles and Starlink customer support.
The idea is simple: developers no longer need to build systems from scratch to transcribe calls, create voice assistants or generate narration. They can connect these APIs to an application, a customer service platform, an accessibility tool or an interactive podcast.
More context in transcription
Grok Speech to Text does more than return a list of words. It can include word-level timestamps, identify speakers and separate audio channels. This makes it possible, for example, to automatically distinguish between a customer and an agent in a recorded call.
It also includes inverse text normalization, a feature that turns spoken language into a more useful written version. Instead of transcribing “one hundred euros on the fifteenth of May” word for word, it can format it as “100 euros, May 15,” depending on the context.
The API supports more than 25 languages and allows the language to change during a conversation. It is designed both for prerecorded files and real-time transcription, which is useful for meetings, captions and customer service centers.
xAI compares Grok STT's performance with commercial models from ElevenLabs, Deepgram and AssemblyAI. The metric used is word error rate, or WER: the lower the rate, the fewer errors the transcription contains.
| Audio type | Grok STT | ElevenLabs | Deepgram | AssemblyAI |
|---|---|---|---|---|
| Call entities | 5.0% | 12.0% | 13.5% | 21.3% |
| Videos and podcasts | 2.4% | 2.4% | 3.0% | 3.2% |
| Meetings | 10.9% | 12.2% | 16.3% | 15.7% |
| Telephony | 9.3% | 9.4% | 11.0% | 11.2% |
| Overall result | 6.9% | 9.0% | 11.0% | 12.9% |
These results come from evaluations published by xAI, not from an independent test. Even so, they point to one of the uses the company highlights: recognizing names, terms and important data in fields such as healthcare, law and finance.
Voices controlled with direct instructions
Grok Text to Speech can generate voices with changes in pacing and expression through simple tags. A developer can add instructions such as [laugh], [sigh], [whisper], <emphasis>, <slow> or <pause> to control how a sentence is delivered.
In practice, this makes it possible to create an assistant that pauses before an important answer, reads a news story with more emphasis or adapts its tone in an educational experience. There is no need to design a complex audio markup system.
Pricing and what changes for you
Transcription costs $0.10 per hour for batch processing and $0.20 per hour for streaming, meaning the audio is processed as it happens. Voice generation costs $15 per million characters.
For users, the change is not a new voice application to install. It is that more services can add conversation, captions or automated reading without building the entire infrastructure themselves. What to watch now is real-world performance across different accents, languages and noisy environments, along with the usage limits xAI sets in its API console.