Grok STT Review: Transcription, Word Timings and No Speaker Labels
One audio file per request, 1 second to 2 hours.
xAI's Grok speech-to-text, generally available since April 2026: send an audio URL and get the transcript back as text. 25 languages with automatic detection, separate transcripts per channel, key-term hints and numbers in written form. Billed per second of audio, and SpicyAPI adds no content filter.
- Lydia
- Tested Sep 30, 2026
- Updated Oct 7, 2026
The verdict
We did not create new paid samples for this article. Published examples remain examples of a particular input and output, rather than a statistically controlled comparison. The illustration accompanying this article is editorial artwork, not a sample produced by the reviewed model.
What we liked and what held it back
What we liked
- Checking generated speech. Run a text-to-speech result back through it and compare the transcript with the script: skipped words, a truncated ending or a tag read out loud show up at once.
- Names, brands and numbers. Key terms let you supply unusual names, while text normalization controls written forms for numbers and dates. Check the returned transcript against the recording; neither control guarantees correct recognition.
What held it back
- One audio file per request, 1 second to 2 hours. Word timings and the detected language come back in the structured transcript, but speaker labels do not, and speakers sharing one channel are not told apart. Raw header-less audio (PCM, μ-law, A-law) is not accepted — send a file in a container such as MP3, WAV, M4A, FLAC or OGG.
What the evidence shows
Our existing 30 September 2026 diagnostic used six LibriSpeech test-clean clips, six speakers, 53.6 seconds of English read speech and 146 reference words. Grok STT recorded 1.37% word error rate: 2 substitutions, 0 deletions and 0 insertions. The transcription methodology explains normalization and the limited task being measured. These are documented earlier results, reused here without a new transcription request.
This small, clean, short-form sample does not establish accuracy on meetings, overlapping speakers, noise, other languages or sung lyrics. Word error rate measures word differences against a reference; it does not measure punctuation, formatting or speaker attribution. Keep that distinction when choosing a tool for subtitles, a searchable archive or a final transcript.
A practical starting brief
Use an audio file that you own or have permission to process. Check the beginning and ending, one proper name and a difficult passage against the audio. For a noisy recording, keep a written record of words you cannot verify instead of silently treating the transcript as authoritative. The model page explains the supported file input.
Start with the material that determines acceptance: a name, a difficult number, a sentence with an emotional turn, or the passage you cannot afford to misread. Listen or read against the original, including the ending. Keep the input fixed while changing one control, so a difference in wording does not get mistaken for a better model. A pleasant voice alone does not prove that every word survived.
Read the current quote on the model task page before submitting. Per-character, per-second and per-request billing cannot be compared by the displayed starting number alone. Compare the same passage or recording, and include any format conversion or alignment you still need to do.
How to choose an alternative
Choose by the missing capability before comparing a headline quality number. An alternative should remove a concrete bottleneck—voice selection, a longer script, delivery direction or the right kind of transcription—rather than simply have a newer name. The current pages below show their actual tasks and price units.
What Grok STT can do on SpicyAPI
Straight from the live model specification on SpicyAPI. Some creators offer more in their own apps; this is what you can call through our API today.
| Feature | Speech to text |
|---|---|
| Seed control | No |
| Negative prompt | No |
What do you think?
Was this review useful?
Log in to vote.
Comments
Comments are checked automatically and by our team. We remove spam, offensive content, personal attacks and anything that names a real person without consent. Report a comment and we review it within one business day.
Tried this model? Share what worked. You need a free account to comment, so we can keep spam out.
Grok STT FAQ
Try Grok STT yourself
Set every parameter in the Playground and see the exact price before you run it.
More reviews
See the leaderboards →
HeartMuLa Transcribe
Lyric text without timestamps, confidence values or speaker labels.
Tested Sep 30, 2026
Read the review
Gemini 3.8 Flash-Lite TTS
One or two voices per request: up to 5,000 characters of script for one voice, or up to 20 dialogue lines of 250 characters each for two.
Not hands-on tested · catalogue checked Oct 7, 2026
Read the review
Gemini 3.8 Flash TTS
One or two voices per request: up to 5,000 characters of script for one voice, or up to 20 dialogue lines of 250 characters each for two.
Not hands-on tested · catalogue checked Oct 7, 2026
Read the review