Speech to Text

Knox exposes ElevenLabs-compatible speech-to-text endpoints at https://api.knoxstudio.ai/v1 for KnoxStudio. Authenticate with your Knox API key (Authorization: Bearer sk-... or xi-api-key). The same routes are also served by https://api.knox.chat/v1.

This is not sending input_audio to chat completions, and it is not Realtime speech-to-speech. STT turns audio or video into text.

Typical flow:

1. List models — GET /v1/speech-to-text/models
2. Transcribe — POST /v1/speech-to-text
3. Optionally fetch or delete — GET / DELETE /v1/speech-to-text/transcripts/{transcription_id}

https://api.knoxstudio.ai/v1
Authorization: Bearer sk-...

Models and billing

curl https://api.knoxstudio.ai/v1/speech-to-text/models

model_id	Use	Price
scribe_v2	Default batch transcription, 90+ languages	$0.27 / audio hour
scribe_v2_realtime	WebSocket streaming, ~150ms	$0.44 / audio hour

Add-ons (hourly, on top of the model rate):

Add-on	Price
Entity detection / redaction	$0.12 / hour
Keyterm prompting	$0.10 / hour
Speaker role detection	+10% of the base transcription cost

When use_multi_channel is true, each channel is billed independently at the full audio duration. More than 100 keyterms also applies a 20 second minimum billable duration.

Multipart uploads may be up to 5GB. Audio must be at least 100ms.

Create a transcript

Provide exactly one of file or source_url (cloud_storage_url is deprecated). model_id is required.

cURL

curl -X POST https://api.knoxstudio.ai/v1/speech-to-text \
     -H "Authorization: Bearer $KNOXCHAT_API_KEY" \
     -F "model_id=scribe_v2" \
     -F "file=@interview.mp3" \
     -F "diarize=true" \
     -F "timestamps_granularity=word"

Python

import os
import requests

with open("interview.mp3", "rb") as audio:
    response = requests.post(
        "https://api.knoxstudio.ai/v1/speech-to-text",
        headers={"Authorization": f"Bearer {os.environ['KNOXCHAT_API_KEY']}"},
        files={"file": audio},
        data={
            "model_id": "scribe_v2",
            "diarize": "true",
            "timestamps_granularity": "word",
        },
    )
response.raise_for_status()
print(response.json()["text"])

A synchronous 200 includes text, language_code, language_probability, words, and optional transcription_id. Set webhook=true to return early and receive the result later.

Useful fields: language_code, tag_audio_events, num_speakers (max 32), diarize, timestamps_granularity (none / word / character), no_verbatim (Scribe v2), keyterms (max 1000).

Realtime WebSocket

GET / POST /v1/speech-to-text/realtime requires a WebSocket upgrade. Default model is scribe_v2_realtime.

Send:

{
  "message_type": "input_audio_chunk",
  "audio_base_64": "<base64 pcm>",
  "commit": false,
  "sample_rate": 16000
}

Receive partial_transcript (interim) and committed_transcript (stable). commit_strategy is manual or vad. Audio formats: pcm_8000, pcm_16000, pcm_22050, pcm_24000, pcm_44100, pcm_48000, ulaw_8000.

Do not put credentials in WebSocket JSON bodies. Knox strips xi-api-key from client messages.

Errors

422 uses { "detail": [{ "loc": [...], "msg": "...", "type": "value_error" }] }. Other errors use { "detail": "..." }.

Common statuses: 401 missing/invalid key, 402 insufficient credits, 403 model not allowed, 400 HTTP to the realtime path without Upgrade, 422 missing model_id / file / URL.