Copy page

Docs

OpenAI SDK clients

OpenAI SDK clients

Soynade exposes a separate OpenAI-compatible surface for developers who want familiar SDK clients and chat.completions response shapes while keeping Soynade authentication, billing, quota enforcement, project attribution, and observability.

Oolel Speech V1 is designed for text translation, speech recognition, speech translation, and streaming translation. Use the dedicated Soynade task endpoints when you want multipart file upload, async jobs, batch processing, or capability-shaped responses.

Shared Setup

Configure OpenAI SDK clients with Soynade's OpenAI-compatible base URL:

Code
https://api.soynade.ai/openai/v1

Use your Soynade API key as the SDK API key:

Code
Authorization: Bearer <your-soynade-api-key>

The surface publishes these endpoints:

Endpoint	Purpose
GET /openai/v1/models	List Soynade models available through the OpenAI-compatible surface.
POST /openai/v1/chat/completions	Run a task-focused completion with JSON or SSE streaming responses.

Current model:

Code
oolel-speech-v1

The supported wire protocol is chat.completions. This does not add native OpenAI audio endpoints.

Python Setup

Install the OpenAI Python SDK:

Code
pip install openai

Create one shared client:

Code
import os
from openai import OpenAI

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

Choose The Right Surface

Use case	Recommended API
Capability-shaped text translation	POST /v1/translations
Multipart audio transcription	POST /v1/audio/transcriptions
Multipart speech translation	POST /v1/audio/translations
OpenAI SDK text translation	POST /openai/v1/chat/completions
OpenAI SDK hosted-audio recognition or translation	POST /openai/v1/chat/completions
Streaming translation	POST /openai/v1/chat/completions

The dedicated task routes are capability-first and do not accept a model selector. The OpenAI-compatible surface is separate and requires the public model id.

Text Translation

For a translation-shaped response and capability-specific metering, use the separate POST /v1/translations endpoint.

Speech Recognition

For speech recognition, use temperature: 0 so transcription is deterministic.

Start with a local audio file by encoding it as base64 and sending it through an input_audio content part.

You can also send hosted audio through an audio_url content part.

For local multipart uploads or asynchronous jobs, use the separate audio transcription endpoints.

Speech Translation

For local multipart uploads or asynchronous jobs, use the separate speech translation endpoints.

Streaming Translation

Set stream=True with the Python client to receive incremental translation chunks:

Audio Input Notes

The OpenAI-compatible surface accepts hosted audio URLs or small local files encoded as data: URLs. It does not accept multipart file uploads. For larger local files or production file-upload workflows, use:

Code
POST /v1/audio/transcriptions
POST /v1/audio/translations

The OpenAI-compatible JSON request is limited to 70,000,000 decimal bytes. All hosted, data-URL, and inline audio sources in one request share a 50,000,000-byte decoded-audio limit. See Errors and limits for the complete table.

Usage And Billing

OpenAI-compatible requests are attributed to the API key and project that made the request. Usage appears in the Soynade dashboard as Oolel Speech V1 usage.

Billing is split into three meters:

Meter	Unit
Oolel Speech V1 input	Input tokens
Oolel Speech V1 output	Output tokens
Oolel Speech V1 audio	Audio prompt seconds

Dedicated product routes keep their capability-specific pricing:

Product route	Meter
Text translation	Text translation characters
Audio transcription	ASR audio seconds
Speech translation	Speech translation audio seconds
Text to speech	TTS characters

Compatibility Notes

• Use oolel-speech-v1 as the model id.
• Continue using client.chat.completions.create(...) for every example on this page.
• Use modalities: ["text"] in these task requests.
• stream: true returns text/event-stream.
• Unsupported model ids return model_not_found.
• Missing credits return a billing error before inference starts.