Inference Logoinference shell

Grok Voice

Talk with Grok. Stream microphone audio over a socket and hear xAI's speech-to-speech model answer as it speaks, with both sides transcribed; the conversation is the task's output.

speech to speechrealtime
run in browser →run via APIrun with your agent
maker
xai
category
audio
input → output
speech → speech
runs via
api, javascript and python sdk, belt cli, mcp
updated
2026-09-24
price
$0.08/min session audio + $0.004/text input

api reference

about

talk with grok. stream microphone audio over a socket and hear xai's speech-to-speech model answer as it speaks, with both sides transcribed; the conversation is the task's output.

get an api key

create one in settings → api keys. the examples read it from INFERENCE_API_KEY.

bash
1export INFERENCE_API_KEY="inf_your_key"

1. calling the api

install the client

nothing to install. any http client works; the examples use curl, and jq to read responses. every response wraps its result in a data field.

run and get result

send "wait": true to hold the connection until the task finishes. without it the call returns right away with the task still running.

bash
12# "wait": true holds the connection until the task finishes.3# the response is {"data": {"id", "status", "status_text", "output", ...}}4curl -X POST https://api.inference.sh/apps/run \5  -H "Authorization: Bearer $INFERENCE_API_KEY" \6  -H "Content-Type: application/json" \7  -d '{8    "app": "xai/grok-voice",9    "input": {10      "events": [11        {12          "text": "text"13        }14      ]15    },16    "wait": true17  }'

stream live updates

get real-time progress updates as the task runs. ideal for showing progress bars, partial results, or long-running tasks. send "stream": true and the response is a server-sent event stream that ends when the task does.

bash
12# "stream": true answers with server-sent events until the task finishes3curl -N -X POST https://api.inference.sh/apps/run \4  -H "Authorization: Bearer $INFERENCE_API_KEY" \5  -H "Content-Type: application/json" \6  -d '{7    "app": "xai/grok-voice",8    "input": {9      "events": [10        {11          "text": "text"12        }13      ]14    },15    "stream": true16  }'

run in the background

start the task and get its id back right away, then check on it later. for long tasks, or when you'd rather be called back, see webhooks.

bash
1# without "wait" the call returns the task right away (status 1, "received")2TASK_ID=$(curl -s -X POST https://api.inference.sh/apps/run \3  -H "Authorization: Bearer $INFERENCE_API_KEY" \4  -H "Content-Type: application/json" \5  -d '{6    "app": "xai/grok-voice",7    "input": {8      "events": [9        {10          "text": "text"11        }12      ]13    }14  }' | jq -r '.data.id')1516# later: read its state. done when status is 10 (completed), 11 (failed) or 12 (cancelled)17curl -s https://api.inference.sh/tasks/$TASK_ID \18  -H "Authorization: Bearer $INFERENCE_API_KEY" | jq '.data | {status, status_text, output}'

2. authentication

the api uses api keys for authentication. see the authentication docs for detailed setup instructions.

3. files

file inputs are automatically handled by the sdk. you can pass local paths, urls, or base64 data.

automatic upload

pass urls directly or use data uris for inline content.

bash
1# pass URLs or data URIs as input values2curl -X POST https://api.inference.sh/apps/run \3  -H "Authorization: Bearer $INFERENCE_API_KEY" \4  -H "Content-Type: application/json" \5  -d '{6    "app": "xai/grok-voice",7    "input": {8      "image": "https://example.com/image.png",9      "audio": "data:audio/mp3;base64,..."10    },11    "wait": true12  }'

manual upload

you can also upload files manually and use the returned url.

bash
1# 1. create the file: the response has an upload_url and the file's uri2FILE=$(curl -s -X POST https://api.inference.sh/files \3  -H "Authorization: Bearer $INFERENCE_API_KEY" \4  -H "Content-Type: application/json" \5  -d '{"files": [{"filename": "image.png", "content_type": "image/png"}]}')67# 2. PUT the bytes to upload_url8curl -X PUT "$(echo "$FILE" | jq -r '.data[0].upload_url')" \9  -H "Content-Type: image/png" \10  --data-binary @image.png1112# 3. pass the uri as an input value13echo "$FILE" | jq -r '.data[0].uri'

4. webhooks

get notified when a task completes by providing a webhook url. when the task reaches a terminal state (completed, failed, or cancelled), a POST request is sent to your url with the task result.

bash
1curl -X POST https://api.inference.sh/apps/run \2  -H "Authorization: Bearer $INFERENCE_API_KEY" \3  -H "Content-Type: application/json" \4  -d '{5    "app": "xai/grok-voice",6    "input": {7        "events": [8            {9                "text": "text"10            }11        ]12    },13    "webhook": "https://your-server.com/webhook"14  }'

webhook payload

your endpoint receives a JSON POST with the task result:

json
1{2  "event": "task.completed",3  "timestamp": "2024-01-15T10:30:05Z",4  "data": {5    "id": "task_abc123",6    "short_id": "abc123",7    "status": 10,8    "status_text": "completed",9    "output": { ... },10    "created_at": "2024-01-15T10:30:00Z",11    "updated_at": "2024-01-15T10:30:05Z"12  }13}
eventstringtask.completed, task.failed or task.cancelled
timestampstringiso timestamp of the delivery
data.idstringtask id
data.short_idstringshort task id
data.statusnumberterminal status (10=completed, 11=failed, 12=cancelled, 13=rejected)
data.status_textstringthe status as a word: completed, failed or cancelled
data.outputobjecttask output (when completed)
data.errorstringerror message (when failed, omitted otherwise)
data.session_idstringsession id (when the task ran in a session, omitted otherwise)
data.created_atstringiso timestamp
data.updated_atstringiso timestamp

5. schema

input

audioarray(stream)

microphone audio, a frame every 20 ms or so

eventsarray(stream)

typed messages: text from the user, or words for the assistant

instructionsstring

system prompt. grok's voice models take plain instructions; workarounds written for other models are unnecessary.

default: "You are a helpful assistant."
voicestring

a built-in voice; `voices` lists them with what xai says about each

default: "eve"
options:"eve""ara""rex""sal""leo""altair""atlas""aurora""carina""castor""celeste""cosmo""helios""helix""iris""kepler""liora""lumen""luna""lux""naksh""orion""perseus""rigel""sirius""ursa""zagan""zenith"
custom_voicestring

the id of a voice cloned with xai's custom voices api. set, it is used instead of voice.

modelstring

grok-voice-latest follows the newest model; pin a versioned name such as grok-voice-think-fast-2.0 for stability

default: "grok-voice-latest"
reasoningstring

whether the model thinks before it answers. 'none' answers faster.

default: "high"
options:"high""none"
languagestring

bcp-47 hint for what the user speaks, such as en, ja or es-mx (spanish and portuguese need a region). left empty it is detected.

speednumber

playback speed of the assistant's voice

default: 1min:0.7max:1.5
idle_minutesnumber

end the session after this long with nobody speaking or typing (0: never). grok bills every minute the session is open, silent or not.

default: 2min:0max:60
silence_msinteger

silence that ends the user's turn, in ms. left empty grok decides.

min:0max:10000
web_searchboolean

let the assistant search the web

default: false
x_searchboolean

let the assistant search x

default: false

output

assistant_textstring

what the assistant is saying, as it says it

audioarray(stream)

the assistant's voice

end_reasonstring

why the session ended, on the result: the caller closed it, or grok did (for example after 15 minutes without audio)

messagesarray

the conversation so far, a message per turn

partialboolean

true while the conversation goes on; false on the result

secondsnumber

how long the session has run

user_textstring

what grok hears the user saying, refined as they speak

other audio apps

Fish Audio Drama 3

fish-audio/drama-3

$50.00/1M UTF-8 bytes

Eleven Flash v2.5

elevenlabs/eleven-flash-v2-5

$0.04/1K characters

Eleven Multilingual v2

elevenlabs/eleven-multilingual-v2

$0.08/1K characters

Eleven v3

elevenlabs/eleven-v3

$0.08/1K characters

Eleven v4 Turbo

elevenlabs/eleven-v4-turbo

$0.04/1K characters

more from xai

Grok Speech to Text

xai/grok-stt

$0.10/hr of audio for files, $0.20/hr for realtime

Grok Imagine Video 1.5

xai/grok-imagine-video-1-5

$0.08/sec 480p, $0.14/sec 720p, $0.25/sec 1080p; $0.01/input image

Grok Imagine Image 2.0

xai/grok-imagine-image-2-0

$0.04/img low 1k, $0.06/img low 2k, $0.06/img med 1k, $0.08/img med 2k; $0.01/source img (edits)

Grok 4.7

xai/grok-4-7

$2.00/M input, $0.50/M cached, $6.00/M output (2x at 200k+ prompt tokens)

new models

  1. J

    JEV-9B

    autotrust/jev-9b

    decisionfrom $0.00055833/sec
  2. A

    AutoTrust-JEV-27B v1.1

    autotrust/jev-27b-v1-1

    decisionfrom $0.00055833/sec

  3. Gemini Omni 1.2 Flash

    google/gemini-omni-1-2-flash

    video$0.0232/sec 360p, $0.0695/sec 720p, $0.1043/sec 1080p, $0.2085/sec 4K, plus $1.50/M input and $9.00/M thinking tokens

  4. GPT Image 2.6 Flare

    openai/gpt-image-2-6-flare

    image$30.00/M output, $5.00/M text input, $8.00/M image input, $0.10/M SVG input, $0.50/M SVG output

  5. GPT Image 2.6 Sunburst

    openai/gpt-image-2-6-sunburst

    image$30.00/M out, $5.00/M text in, $8.00/M img in; SVG $0.10/$0.50/M in/out

ready to run Grok Voice?

try in browserbrowse all apps

we use cookies

we use cookies to ensure you get the best experience on our website. for more information on how we use cookies, please see our cookie policy.

by clicking "accept", you agree to our use of cookies.
learn more.