Build with audio.

Speech models price per character (pricing.speech_character in the catalog, shown per 1M characters). Transcription models price per second (pricing.transcription_second, shown per minute). Both list prompt and completion as 0. Prices include the MiniRouter fee.

Speech holds are exact: characters × the per-character rate. Transcription holds read the duration from the uploaded file and reserve it, rounded up to a second, × the per-second rate. Audio contributes 32 prompt tokens per second to the hold, alongside other input and reserved output tokens. Each hold settles to the provider's reported usage and the rest is released. Request logs rows show speech_characters or transcription_seconds with zero tokens; the CSV export carries both columns.

Body: model, input (up to 4,096 characters), optional voice, response_format (mp3, wav, opus, pcm, aac, flac; mp3 by default), speed and instructions. The response is the audio file with that format's Content-Type. No streaming.

Voices are per provider: OpenAI alloy, ash, coral, echo, fable, nova, onyx, sage, shimmer; Microsoft en-US-Harper, en-US-Nova, en-US-Davis, en-US-Jenny. An unknown voice returns 400.

// npm install openai@7.4.0
import OpenAI from 'openai';
import { writeFile } from 'node:fs/promises';

const client = new OpenAI({
  baseURL: 'https://api.minirouter.sh/v1',
  apiKey: process.env.MINIROUTER_API_KEY,
  maxRetries: 0,
});
const response = await client.audio.speech.create({
  model: 'openai/tts-1',
  input: 'The quick brown fox jumps over the lazy dog.',
  voice: 'alloy',
  response_format: 'mp3',
});
await writeFile('speech.mp3', Buffer.from(await response.arrayBuffer()));
console.log(response.headers.get('x-minirouter-request-id'));

# pip install openai==2.53.0
import os
from openai import OpenAI

client = OpenAI(
    base_url='https://api.minirouter.sh/v1',
    api_key=os.environ['MINIROUTER_API_KEY'],
    max_retries=0,
)
raw = client.audio.speech.with_raw_response.create(
    model='openai/tts-1',
    input='The quick brown fox jumps over the lazy dog.',
    voice='alloy',
    response_format='mp3',
)
with open('speech.mp3', 'wb') as output:
    output.write(raw.content)
print(raw.headers.get('x-minirouter-request-id'))

Multipart: file, model, optional language (accepted but currently ignored upstream) and response_format (json default, text, verbose_json). Uploads are capped at 8 MiB.

JSON body: model, audio (base64), media_type, optional language and response_format. The same 8 MiB cap applies to the decoded audio, and a larger body is refused with 413. Write the body to a file so large recordings do not hit the shell's argument limit.

{
  printf '{"model":"openai/whisper-1","media_type":"audio/wav","response_format":"json","audio":"'
  base64 < meeting.wav | tr -d '\n'
  printf '"}'
} > body.json

curl --fail-with-body --retry 0 --max-time 120 \
  https://api.minirouter.sh/v1/audio/transcriptions \
  -H "Authorization: Bearer $MINIROUTER_API_KEY" \
  -H 'Content-Type: application/json' \
  --data-binary @body.json

json returns { "text": "…" }. text returns the transcript as plain text. verbose_json returns:

{
  "text": "The quick brown fox jumps over the lazy dog.",
  "language": "en",
  "duration": 2.49,
  "segments": [{ "start": 0, "end": 2.49, "text": "The quick brown fox jumps over the lazy dog." }]
}

language and duration can be null when the provider does not report them. segments can be empty; segment timestamps are included only when available.

Accepted containers: WAV, FLAC, OGG, MP4/M4A, MP3, up to 3,600 s; WebM is rejected. Microsoft transcription also rejects MP4/M4A.

// npm install openai@7.4.0
import OpenAI from 'openai';
import { createReadStream } from 'node:fs';

const client = new OpenAI({
  baseURL: 'https://api.minirouter.sh/v1',
  apiKey: process.env.MINIROUTER_API_KEY,
  maxRetries: 0,
});
const result = await client.audio.transcriptions.create({
  model: 'openai/whisper-1',
  file: createReadStream('meeting.wav'),
  response_format: 'verbose_json',
}).withResponse();
console.log(result.data.text, result.data.duration);
console.log(result.response.headers.get('x-minirouter-request-id'));

# pip install openai==2.53.0
import os
from openai import OpenAI

client = OpenAI(
    base_url='https://api.minirouter.sh/v1',
    api_key=os.environ['MINIROUTER_API_KEY'],
    max_retries=0,
)
with open('meeting.wav', 'rb') as audio:
    raw = client.audio.transcriptions.with_raw_response.create(
        model='openai/whisper-1',
        file=audio,
        response_format='verbose_json',
    )
print(raw.parse().text)
print(raw.headers.get('x-minirouter-request-id'))

Send a recording to a language model as a file content part whose file_data is an audio/* data URL. Only models whose catalog input_modalities include audio accept it, for example google/gemini-3.6-flash and xiaomi/mimo-v2.6-flash. Containers: WAV, FLAC, OGG, MP4/M4A, MP3, up to 3,600 s per request. Audio is billed as ordinary prompt tokens at the model's input rate.

On /v1/responses use an input_file part:

{
  printf '{"model":"xiaomi/mimo-v2.6-flash","max_output_tokens":256,"input":[{"role":"user","content":['
  printf '{"type":"input_text","text":"Summarize this recording in one sentence."},'
  printf '{"type":"input_file","filename":"clip.mp3","file_data":"data:audio/mpeg;base64,'
  base64 < clip.mp3 | tr -d '\n'
  printf '"}]}]}'
} > body.json

curl --fail-with-body --retry 0 --max-time 120 \
  https://api.minirouter.sh/v1/responses \
  -H "Authorization: Bearer $MINIROUTER_API_KEY" \
  -H 'Content-Type: application/json' \
  --data-binary @body.json

input_audio parts are rejected with 400.

Speech and transcription models are listed in the voice catalog with type speech or transcription. Language models that accept audio list audio in architecture.input_modalities. Your API key must permit the model. Unavailable routes return HTTP 503 before a credit hold or provider call.

Use a MiniRouter API key as a bearer token. These examples name a concrete model that accepts the input.

Speech and transcription return complete files or text. These two endpoints do not support streaming, realtime, saved presets, Auto, Fusion or fallback model lists.

Chat and Responses accept inline audio files on eligible models. Saved presets and supported model fallbacks follow the normal routing rules; every selected route must accept audio. input_audio parts, the chat audio field and audio output in modalities are rejected. Audio is not supported with provider keys (BYOK).

Disable automatic retries: a lost response may still incur cost. Keep x-minirouter-request-id and read the settled charge in Request logs.

• Model catalog
• MiniRouter authentication
• MiniRouter billing
• Errors