Text to speech

POST /v1/audio/speech follows the OpenAI speech API: send text, get audio bytes back. Speech models are listed under Models → Audio. Some are priced per second of audio, others per 10,000 characters of input.

#List voices

curl https://bigmodel.io/v1/audio/voices \
  -H "Authorization: Bearer $BIGMODEL_API_KEY"

{"object": "list", "data": [{"id": "...", "name": "...", "language": "zh", "tags": ["female", "calm"]}]}

Pass a voice's id as voice. Without voice, the model's default voice is used.

#Generate speech

curl https://bigmodel.io/v1/audio/speech \
  -H "Authorization: Bearer $BIGMODEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "minimax-speech-2.8-turbo", "input": "Welcome to BigModel.io. One key for every model.", "voice": "VOICE_ID"}' \
  --output speech.mp3

import os
from openai import OpenAI

client = OpenAI(base_url="https://bigmodel.io/v1", api_key=os.environ["BIGMODEL_API_KEY"])

with client.audio.speech.with_streaming_response.create(
    model="minimax-speech-2.8-turbo",
    voice="VOICE_ID",
    input="Welcome to BigModel.io. One key for every model.",
) as response:
    response.stream_to_file("speech.mp3")

The response body is the audio file, with its type in Content-Type (usually audio/mpeg) and the task ID in X-Task-Id.

#Long texts

The request waits up to 2 minutes. If the audio is not ready by then, you get 202 with a task (object: "audio.generation") and a Location header instead of audio. Poll GET /v1/audio/generations/{id} and download from GET /v1/audio/generations/{id}/content, as described in Music, sound effects & audio tools. To always work asynchronously, send the same body to POST /v1/audio/generations.

#How you are charged

Per-10k-character models charge for the characters in input. Per-second models charge for the length of the audio produced. Add "dry_run": true to the body to get {"object": "quote", "estimated_cost_usd": ...} without generating anything. Failed requests are not charged. Charges appear in Usage.

#Verify the setup

Call /v1/audio/voices, then run the generation example with one of the IDs. A playable speech.mp3 means the key works.

#Troubleshooting

400 "input is required" or "input must be at most 4000 characters". Send text in input and split long texts into several requests.

400 "speed must be between 0.5 and 2". Use a value in that range.

404 model_not_available. The model is not offered to this key, or it is not a speech model. Check GET /v1/models.

402 insufficient_balance. Top up in Billing.

502 upstream_error. The speech could not be generated. Try a different voice or text. You are not charged.

data is empty in /v1/audio/voices. No speech model is offered to this key's group.

#Related

• Video generation: avatar models use a voice track as input_audio.
• Music, sound effects & audio tools.
• Pricing for speech prices.