Who makes Grok STT and when was it released?

xAI makes it, and it was released on 15 April 2026. On SpicyAPI it is a single-endpoint family: one task, speech to text, called as xai/grok-stt/speech-to-text. It takes an audio URL and returns text, so nothing is written to media storage.

Can it tell speakers apart?

Not by voice: speaker labels are not returned. If each speaker was recorded on their own channel, as in many call recordings, turn on multichannel and each channel comes back on its own line. Without it, a stereo file with a different speaker on each side lost one of them in both of our tests.

Why is more held than I am charged?

The length of the audio is only known once it has been read, so the price of the two-hour maximum is held when the task is created. You are charged for the actual seconds of audio, and the rest of the hold is freed when the task finishes. None of the options changes the price, and a failed request is not charged.

Why does text_normalization need a language?

Numbers, currencies and dates are written differently in each language, so the model has to know which convention to follow. Without language the request is rejected before it runs. When you do not need digits, leave language empty and it is detected for you.

Should I use this or HeartMuLa Transcribe?

It depends on what is in the audio. If you are transcribing people talking, such as calls, meetings, interviews or spoken-word audio, and you want word timings, one line per channel or a list of names to get right, use Grok STT. If you are pulling lyrics out of finished songs, look at HeartMuLa Transcribe, which is aimed at sung vocals.

Grok STT API — Speech to Text, Unfiltered Transcripts

xai/grok-stt/speech-to-text

Grok STT is xAI's speech-to-text: send an audio URL and get back the transcript as text plus word-level timings and the detected language. It covers 25 languages, with per-channel transcripts and key-term hints.

From $0.00003/s

Run model View API

Input

Speech to text

Input audio1/1

Drop an audio file here, or browse · 1 file · Max 90 MB · MP4 · MP3 · WAV

Language

Normalize text

Key terms0/100

Keep filler words

Transcribe channels separately

Checking your account before enabling submission and pricing.

Output

Ready for a new run

Your result will appear here.

Configure the inputs, review the estimate, then run the model.

{
  "model": "xai/grok-stt/speech-to-text",
  "input": {
    "text_normalization": false,
    "filler_words": false,
    "multichannel": true,
    "audio_url": "https://cdn.spicyapi.ai/models/examples/grok-stt/a8922505cb831dc4.mp3"
  }
}

Introduction

Send an audio URL and, optionally, the spoken language, a list of key terms and three switches: written-form numbers, filler words and per-channel transcription. The task returns the transcript as text plus a structured transcript with word-level timings, the detected language and the audio length; no file is produced. SpicyAPI applies no content filter: unfiltered speech is transcribed as spoken. You are billed per second of audio.

Specifications for Grok STT · Speech to Text
Field	Value
Input	Audio
Output	Text
Required field	`audio_url` — an uploaded file or a public audio URL, 1 second to 2 hours
Optional fields	language (25 codes), `text_normalization`, keyterms (up to 100), `filler_words`, multichannel
Result	Transcript as text, plus output.transcript with word timings, language and duration; per channel with multichannel
Billed quantity	Seconds of audio; uploads are held at their length, URLs at the two-hour maximum
Typical time	About 3 seconds for a 10-second clip, about 15 seconds for an hour
Retention	Uploaded audio is deleted after one day; the transcript is erased with prompt data

Request parameters

Request parameters for Grok STT · Speech to Text
Parameter	Required	Description
`audio_url`	Yes	The audio to transcribe: an uploaded MP3 or WAV, or a public URL to MP3, WAV, M4A, FLAC, OGG, Opus, AAC, MP4 or MKV. 1 second to 2 hours. at least 1 characters
`language`	No	The spoken language as a language code. Leave it empty to detect the language automatically; speech in any supported language is transcribed either way. `ar` · `cs` · `da` · `nl` · `en` · `fil` · `fr` · `de` · `hi` · `id` · `it` · `ja` · `ko` · `mk` · `ms` · `fa` · `pl` · `pt` · `ro` · `ru` · `es` · `sv` · `th` · `tr` · `vi`
`text_normalization`	No	When on, spoken numbers, prices, dates and units are written in digits — "$1,250" and "March 14" instead of the words. Needs language. Default: `false`
`keyterms`	No	Up to 100 words or short phrases (50 characters each) to bias the transcript toward — names, brands, jargon. Does not change the price. Default: `[]` at most 100 items
`filler_words`	No	When on, filler words such as um and uh are kept in the transcript; off (default) removes them. Does not change the price. Default: `false`
`multichannel`	No	When on, each channel of a multichannel file (up to 8) is transcribed independently and returned on its own line, in channel order. Default: `false`

Pricing

$0.00003 / second of audio

Billed per second of audio. When the task is created, the hold matches the length of audio you uploaded, or the price of the two-hour maximum for an audio URL; you pay for the actual length and the rest of the hold is freed. A failed request is not charged.

Use cases

Voice notes and audio stories

Unfiltered speech is transcribed as spoken, so voice messages, roleplay recordings and audio stories can be searched, captioned or edited as text.

Call recordings

Recordings with each side on its own channel come back as one line per side when multichannel is on. With it off, our tests returned only one of the two sides, so turn it on for split-channel calls.

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

List unusual names as key terms and turn on text normalization, and the transcript spells the names your way and writes prices and dates as $1,250 or March 14.

Explore key use-case tutorials

Prompt guide & pro tips

Point audio_url at the audio

Upload an MP3 or WAV, or pass a public URL that answers with the audio itself — a player page or share link that returns HTML fails. From 1 second to 2 hours.
Add what you know

Set language if you know it, list names and brand words in keyterms, and turn on text_normalization if you want numbers written as digits. text_normalization needs language.
Read the text from the task

The transcript arrives as text on the task result, and output.transcript lists every word with its start and end time. With multichannel on there is one line per channel, in channel order. Copy it into your own storage if you need it later.

What it returns

The transcript as text on the task result, plus output.transcript: the detected language and language code, the audio length in seconds, and every word with its start and end time in seconds. With multichannel on, output.transcript.channels holds the same for each channel. No file is produced and nothing enters media storage. Speaker labels are not included.

Sending audio

audio_url takes an uploaded MP3 or WAV, or a public URL to MP3, WAV, M4A, FLAC, OGG, Opus, AAC, MP4 or MKV, from 1 second to 2 hours. The URL has to answer with the audio itself. Raw audio without a file header (PCM, μ-law, A-law) is not accepted.

Language and numbers

Leave language empty and the language is detected; speech in any of the 25 supported languages is transcribed either way. Set it and turn on text_normalization to get numbers, prices, dates and units in written form — in our test "one thousand two hundred fifty dollars on March fourteenth" became "$1,250 on March 14". text_normalization needs language.

Key terms and filler words

keyterms takes up to 100 terms of up to 50 characters, one per row. In our test "Xandrelle" and "Qorvessa" came back as "Xandrel" and "Corvessa" without them and exactly right with them. filler_words keeps um and uh, which are removed by default.

Several channels

Turn on multichannel for recordings with each speaker on their own channel. Each channel is transcribed separately and returned on its own line, in channel order. Leave it off only when every channel carries the same sound: in our tests a stereo file with a different speaker on each side came back with one of the two speakers missing.

Billing

Content

SpicyAPI adds no content filter. In our tests unfiltered speech came back word for word, with no asterisks, bleeps or gaps.

SpicyAPI adds no content filter: speech is transcribed as spoken rather than masked or refused. You are responsible for the rights to the audio you send and for the consent of the people recorded.

API quick start

# Generate the idempotency key once; reuse it when retrying after a lost response.
export SPICY_IDEMPOTENCY_KEY="${SPICY_IDEMPOTENCY_KEY:-$(uuidgen)}"
curl -X POST https://api.spicyapi.ai/api/v1/jobs/createTask \
  -H "Authorization: Bearer $SPICY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $SPICY_IDEMPOTENCY_KEY" \
  -d '{
  "model": "xai/grok-stt/speech-to-text",
  "input": {
    "audio_url": "https://example.com/reference.mp3"
  }
}'

# Copy data.taskId from the response into TASK_ID, then poll the status.
TASK_ID="TASK_ID_FROM_RESPONSE"
curl "https://api.spicyapi.ai/api/v1/jobs/recordInfo?taskId=$TASK_ID" \
  -H "Authorization: Bearer $SPICY_API_KEY"

See the API reference for authentication, the complete schema, polling, and error handling.

Related models and next tasks

Missing the model you need?

Tell us what to add and where you would use it. Requests go to the people who decide what gets onboarded next.

Speech to text in 25 languages, with the numbers, names and channels the way you want them

Grok STT from xAI takes an audio URL and returns the words as text on the task, plus a structured transcript with every word timed. Leave the language empty and it is detected; set it and numbers can be written as digits. List the names that matter and they come back spelled your way, and switch on multichannel so each channel of a call gets its own line. Nothing is written to media storage, and you are billed for the seconds of audio you actually send.

Open the playground

Input: audio_url: an uploaded MP3 or WAV, or a public URL to a common audio or video format
Length: 1 second to 2 hours per request
Languages: 25 spoken languages; detected automatically when you leave language empty
Output: Transcript text, plus a word-timed transcript; one entry per channel with multichannel
Billing: Per second of audio; none of the options changes the price
Retention: No media asset written; uploaded audio deleted after 1 day

Languages: 25 spoken languages, detected or declaredLeave language empty and the language is detected; speech in any of the 25 supported languages is transcribed either way. Set it when you want text_normalization, which writes numbers the way that language writes them and therefore needs to know which one to use.
Structure: The text on the task, and a timed transcript beside itThe result carries the full transcript as text. Next to it, output.transcript lists the detected language, the audio length in seconds and every word with its start and end time, and with multichannel on it repeats that for each channel. No file is produced. Speaker labels are not part of the result.
Channels: One line per channel when speakers are recorded apartWith multichannel on, each channel of the file (up to 8) is transcribed separately and returned on its own line, in channel order. It is the setting people miss: in our tests a stereo file with a different speaker on each side came back with only one of them when it was off.

What the Grok STT API does well

Numbers written the way people write them
Turn on text_normalization and spoken figures, prices, dates and units come back as digits. In our test, "one thousand two hundred fifty dollars on March fourteenth" became "$1,250 on March 14". It needs language, because every language writes these differently; without it the request is rejected before it runs. Off by default.
Names spelled the way you list them
keyterms takes up to 100 words or short phrases of up to 50 characters, one per row, and biases the transcript toward them. Two invented names in a lounge announcement came back as "Xandrel" and "Corvessa" without the list, and exactly right with it. Use it for guests, brands, product names and jargon.
Both sides of a call, not just one
A recording with the agent on the left channel and the customer on the right needs multichannel. Each side comes back on its own line, and output.transcript.channels carries the word timings for each. Without it, our stereo tests returned only one of the two speakers, and the other side was simply missing.
Um and uh, kept or cleaned
Filler words are removed by default, which gives clean captions and notes. Turn filler_words on for a verbatim record: in our test the transcript then began with "Um, so, uh," exactly as spoken. Neither setting changes the price.
Unfiltered speech comes back as spoken
SpicyAPI adds no content filter. In our tests recordings of our strongest monologue and narration scripts came back word for word, with no asterisks, bleeps or gaps. That keeps search, captions and review working on the same material the rest of your product is built around. The rights to the audio and the consent of the people recorded stay with you.
Long recordings without a long wait
A request takes anything from 1 second to 2 hours of audio. In our runs a 10-second clip came back in about 3 seconds and a full hour in about 15. Submitting returns a task straight away, and the transcript is read from that task once it finishes.

Who Grok STT is for

Podcasters and video editors
Subtitles and show notes from the recording you already have
Every word arrives with a start and end time, so caption lines can be cut to whatever length your player needs. List guest and brand names in keyterms so they are spelled once and correctly, and leave filler_words off for text that reads cleanly.
Support and sales teams
Recorded calls with each side kept apart
Calls are often recorded with one person per channel. With multichannel on, the agent and the customer come back on separate lines, which is what a review, a summary or a search over the call needs. There are no speaker labels, so a recording with everyone mixed into one channel returns a single stream of text.
Meeting and interview notes
Notes where the figures and the names are right
Set language and turn on text_normalization and amounts, dates and quantities arrive as written figures instead of spelled-out words. Add the attendees and product names to keyterms and they stop drifting between spellings from one meeting to the next.
Uncensored audio creators and platforms
A transcription step that does not refuse the material
Audio drama, roleplay recordings and unfiltered audiobook catalogs are transcribed as spoken, so search, captions and review work on the whole library instead of leaving a section with no text. Responsibility for the source material and for the consent of the people in it stays with the caller.

How to call the Grok STT API

Point audio_url at the audio itself
Upload an MP3 or WAV, or pass a public URL that answers with the audio bytes. MP3, WAV, M4A, FLAC, OGG, Opus, AAC, MP4 and MKV all work, from 1 second to 2 hours. A player page or a share link that returns HTML fails, and raw audio without a file header (PCM, μ-law, A-law) is not accepted.
Add what you know about the recording
Set language if you know it, list names and brand words in keyterms, and turn on text_normalization for digits (it needs language). If speakers were recorded on separate channels, turn on multichannel; if you want the ums and uhs, turn on filler_words.
Send it with an Idempotency-Key and wait for the task
Submitting returns a task. Send an Idempotency-Key so that a retry after a dropped connection lands on the same task instead of creating a second one, then poll the task or take the callback you registered.
Read the text, and keep your own copy
The transcript arrives as text on the task result, and output.transcript holds the word timings, the detected language and the audio length. Nothing is written to media storage, and the transcript is erased with the prompt data, so copy it into your own store when the task finishes.

Grok STT worked examples

A stereo call with one speaker on each side
A 12-second recording with a woman on the left channel and a man on the right, sent with multichannel on. Each side came back on its own line. With multichannel off, the same file returned only the woman's words.
```
{"audio_url": "https://files.example.com/calls/support-call-0412.wav", "multichannel": true}
```
Names that have no dictionary spelling
A lounge announcement with two invented names, sent with both listed as key terms. Without the list the same audio came back as "Xandrel" and "Corvessa"; with it the names were exactly as listed.
```
{"audio_url": "https://files.example.com/announcements/lounge-announcement.mp3", "keyterms": [{"term": "Xandrelle"}, {"term": "Qorvessa"}]}
```
A booking confirmation with the figures as digits
A spoken confirmation sent with the language declared and text_normalization on. The price and the date came back as "$1,250" and "March 14" instead of words.
```
{"audio_url": "https://files.example.com/calls/booking-confirmation.mp3", "language": "en", "text_normalization": true}
```

Grok STT API questions

Most of the quality here comes from what you tell it before you run it. Declare the language when you know it, list the names that must not drift, and turn on multichannel the moment a recording keeps its speakers on separate channels; that one setting decides whether half a conversation goes missing. Then read the timed transcript, not only the text, if you are building captions. The price follows the seconds of audio and nothing else, so try it on a short clip first and read the live quote on the page.

Workbench

Input

Output

Introduction

Request parameters

Pricing

Use cases

Voice notes and audio stories

Call recordings

Checking generated speech

Names, brands and numbers

Prompt guide & pro tips

What it returns

Sending audio

Language and numbers

Key terms and filler words

Several channels

Billing

Content

API quick start

Related models and next tasks

Speech to text in 25 languages, with the numbers, names and channels the way you want them

What the Grok STT API does well

Numbers written the way people write them

Names spelled the way you list them

Both sides of a call, not just one

Um and uh, kept or cleaned

Unfiltered speech comes back as spoken

Long recordings without a long wait

Who Grok STT is for

Subtitles and show notes from the recording you already have

Recorded calls with each side kept apart

Notes where the figures and the names are right

A transcription step that does not refuse the material

How to call the Grok STT API

Point audio_url at the audio itself

Add what you know about the recording

Send it with an Idempotency-Key and wait for the task

Read the text, and keep your own copy

Grok STT worked examples

A stereo call with one speaker on each side

Names that have no dictionary spelling

A booking confirmation with the figures as digits

Grok STT API questions

Who makes Grok STT and when was it released?

Can it tell speakers apart?

Why is more held than I am charged?

Why does text_normalization need a language?

Should I use this or HeartMuLa Transcribe?

All Grok models6

Audio models2

Text models4

Related comparisons