Audio (speech and transcription)

Two endpoints in the format of the OpenAI Audio API: one turns text into speech, the other turns a recording into text. The audio you send is processed in memory and never written to our servers. Moon Blur does not apply to the text read aloud or to the audio (X-Moon-Blur: not_applied).

Read a text aloud#Link to the Read a text aloud section

POST https://api.realmoon.ai/v1/audio/speech

• Authentication: an sk-moon-… key with the inference scope (see Authentication).
• Body: JSON, 256 KB at most.
• Response: the audio file itself.

Parameters#Link to the Parameters section

Any other parameter is rejected with 400 unknown_parameter.

Example#Link to the Example section

curl https://api.realmoon.ai/v1/audio/speech \
  -H "Authorization: Bearer $MOONAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "mistralai/voxtral-mini-tts-2603", "voice": "fr_marie_neutral", "input": "Hello and welcome."}' \
  -o hello.mp3

import os
from openai import OpenAI

client = OpenAI(base_url="https://api.realmoon.ai/v1", api_key=os.environ["MOONAI_API_KEY"])

with client.audio.speech.with_streaming_response.create(
    model="mistralai/voxtral-mini-tts-2603",
    voice="fr_marie_neutral",
    input="Hello and welcome.",
) as response:
    print("Credits:", response.headers.get("x-moon-credits"))
    response.stream_to_file("hello.mp3")

import fs from 'node:fs';
import OpenAI from 'openai';

const client = new OpenAI({ baseURL: 'https://api.realmoon.ai/v1', apiKey: process.env.MOONAI_API_KEY });

const speech = await client.audio.speech.create({
  model: 'mistralai/voxtral-mini-tts-2603',
  voice: 'fr_marie_neutral',
  input: 'Hello and welcome.',
});
fs.writeFileSync('hello.mp3', Buffer.from(await speech.arrayBuffer()));

Response#Link to the Response section

The response body is the audio file:

Speech models#Link to the Speech models section

List as of November 5, 2026; the current list, with prices, comes from GET /v1/models?type=speech.

Costs#Link to the Costs section

Most voices are billed per character of text read (billing: per_character). Gemini voices are billed on the text sent and on the length of the audio actually produced. The request holds the price of the text before the call; a request that fails or is cut off by our timeout costs nothing.

Errors#Link to the Errors section

Transcribe a recording#Link to the Transcribe a recording section

POST https://api.realmoon.ai/v1/audio/transcriptions

• Authentication: an sk-moon-… key with the inference scope.
• Body: multipart/form-data only, with the Content-Length header. The file goes in the file field.
• Response: the transcription, as JSON or plain text.

Parameters#Link to the Parameters section

Any other field is rejected with 400 unknown_parameter. A recording longer than 20 minutes is rejected (400 audio_too_long): split it into pieces. For a compressed file, the duration counted is at least what its size implies.

Example#Link to the Example section

curl https://api.realmoon.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $MOONAI_API_KEY" \
  -F model=mistralai/voxtral-mini-transcribe \
  -F language=en \
  -F file=@meeting.m4a

with open("meeting.m4a", "rb") as audio:
    transcript = client.audio.transcriptions.create(
        model="mistralai/voxtral-mini-transcribe",
        file=audio,
        language="en",
    )
print(transcript.text)

const transcript = await client.audio.transcriptions.create({
  model: 'mistralai/voxtral-mini-transcribe',
  file: fs.createReadStream('meeting.m4a'),
  language: 'en',
});
console.log(transcript.text);

Response#Link to the Response section

{
  "text": "Hi everyone, we will start with the budget.",
  "usage": { "type": "duration", "seconds": 312.4, "credits": 0 }
}

The credits value is made up.

With response_format: "text", the response is the text alone, as text/plain. In every case, the X-Moon-Credits header gives the credits charged, and X-Moon-Request-Id the request ID.

Transcription models#Link to the Transcription models section

List as of November 5, 2026; the current list, with prices per minute, comes from GET /v1/models?type=transcription.

• mistralai/voxtral-mini-transcribe, mistralai/voxtral-mini-transcribe-3-0, mistralai/voxtral-small-transcribe-3-0
• openai/whisper-large-v3-turbo, openai/whisper-large-v3, openai/gpt-4o-mini-transcribe, openai/gpt-4o-transcribe
• elevenlabs/scribe-v2, deepgram/nova-3, assemblyai/universal-3-6-pro
• nvidia/parakeet-tdt-0.6b-v3, qwen/qwen3-asr-1.7b

Costs#Link to the Costs section

Transcription is billed at the actual cost reported by the provider, based on the audio duration (or, failing that, the list price). Before the call, the request holds at least the price of one minute, with a margin. A transcription that fails costs nothing.

Errors#Link to the Errors section

Key, balance and spend cap errors, common to both endpoints, are in Errors.

Limits#Link to the Limits section

See also#Link to the See also section

• Models: the default_voice, response_formats and max_file_mb fields.
• Privacy: what Moon AI keeps from audio requests.