# Arabic speech recognition and vision, as an API

- Canonical URL: https://www.broadnet.me/products/inference
- Language: en
- Last updated: 2026-10-21
- Publisher: BroadNet (https://www.broadnet.me)
- Markdown version of the page above, generated from the rendered page.

BroadNet Inference API

Speech recognition, document reading and chat for Arabic and English, on GPUs run by BroadNet. Transcribe every major Arabic dialect, turn invoices and IDs into JSON, and chat over text and images through OpenAI-compatible endpoints.

[Create an account (opens in a new tab)](https://inference.broadnet.ai/) [Developer reference (opens in a new tab)](https://inference.broadnet.ai/docs) [Models and pricing (opens in a new tab)](https://www.broadnet.ai/inference)

New accounts start with $5 of credit, valid for 30 days. No card needed.

**Base URL**

https://inference.broadnet.ai/v1

**Authentication**

Authorization: Bearer <key>

**Arabic transcription**

$0.25 per audio hour

**Content storage**

None: held in memory for a single request

## What it does

- Arabic speech recognition Transcribes Gulf, Egyptian, Levantine, Maghrebi and Modern Standard Arabic. The Arabic model keeps its output in Arabic script, so calls and voice notes come back as searchable Arabic text.
- Documents and vision Photograph an invoice, an ID or a handwritten form and get clean JSON back in Arabic and English, with numbers exactly as written.
- Chat over text and images Gemma 4 VL 26B, tuned and served by BroadNet for Arabic and English, accepts text and images on the chat-completions endpoint and supports streaming.

## Models and pricing

Every model runs on BroadNet hardware. Language models are priced per million tokens and transcription per minute of audio. These are the prices published on broadnet.ai/inference on 24 September 2026; the rate card there is the price of record.

| Model | Model ID | Input type | Price (USD) |
|---|---|---|---|
| Gemma 4 VL 26B | `broadnet/gemma-4-26b-a4b` | Text and images | $0.10 / $0.22 per 1M tokens (input / output) |
| Qwen3-VL 8B Instruct | `broadnet/qwen3-vl-8b` | Text and images | $0.117 / $0.455 per 1M tokens (input / output) |
| Qwen3-VL 30B-A3B Instruct | Private model ID, on request | Text and images | $0.13 / $0.52 per 1M tokens (input / output) |
| Cohere Transcribe Arabic | `broadnet/cohere-transcribe-arabic` | Audio (Arabic) | $0.0042 per minute of audio |
| Cohere Transcribe Arabic fine-tuned by BroadNet | Private model ID, on request | Audio (Arabic) | $0.005 per minute of audio |
| Cohere Transcribe | `broadnet/cohere-transcribe` | Audio (14 languages) | $0.002 per minute of audio |

## Quickstart

Create a key on your account page and send it as a Bearer token. The chat-completions and audio-transcription request shapes are OpenAI-compatible, so the standard SDK works once you point it at the BroadNet base URL.

Transcribe Arabic audio (curl)

```
curl https://inference.broadnet.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $BROADNET_API_KEY" \
  -F model=broadnet/cohere-transcribe-arabic \
  -F language=ar \
  -F file=@voice-note.wav
```

Read an invoice photo (Python)

```
import base64, os
from openai import OpenAI

client = OpenAI(base_url="https://inference.broadnet.ai/v1",
                api_key=os.environ["BROADNET_API_KEY"])

invoice = base64.b64encode(open("invoice.jpg", "rb").read()).decode()
response = client.chat.completions.create(
    model="broadnet/gemma-4-26b-a4b",
    messages=[{"role": "user", "content": [
        {"type": "text", "text": "ما هو المبلغ الإجمالي في هذه الفاتورة؟"},
        {"type": "image_url",
         "image_url": {"url": "data:image/jpeg;base64," + invoice}},
    ]}])
print(response.choices[0].message.content)
```

Always send language with a transcription request; without it, Arabic can be transliterated into Latin letters. Accepted formats are audio/wav and audio/mpeg, up to 26,214,400 bytes per file.

## Where your data goes

Prompts, images, audio, answers and transcripts stay in memory for one request and are then discarded. Content is never logged or used to train or evaluate a model, and an automated audit checks this configuration every hour.

BroadNet keeps only content-free accounting and security records, such as timings, token counts and cost.

Processing runs on hardware BroadNet operates in the United States. Regional processing in MENA or Europe is available on request for data-residency requirements.

## Keys, limits and private endpoints

- Pay-as-you-go keys 30 requests per minute and 4 concurrent requests, paid from a prepaid balance you top up by card. Up to ten live keys per account, each valid for one year.
- Hourly keys $0.50 buys one hour of Gemma 4 VL 26B: one stream, 120 requests per minute and 300,000 tokens per minute, minted by API from your balance.
- IP allow-lists Restrict your keys to your own servers with up to 32 IPv4 or IPv6 addresses or CIDR ranges per account.
- Private endpoints A reserved GPU or node running the model you choose, including Arabic fine-tunes, with your own key, limits and region. Pricing on request.

## Questions

### Which Arabic dialects does the transcription handle?

Gulf, Egyptian, Levantine, Maghrebi and Modern Standard Arabic. Use the Arabic model for Arabic speech: it stays in Arabic script, while the multilingual model can drift into English on mixed speech.

### How do I get an API key?

Create an account at inference.broadnet.ai, add balance, then create or revoke keys from your account page. New accounts start with $5 of credit, valid for 30 days, with no card needed.

### Are my audio files and documents stored?

No. Request and response content is held in memory for one request and then discarded. It is never logged or used for training.

### Where is the data processed?

On BroadNet-operated GPUs in the United States. Regional processing in MENA or Europe is available on request for customers with data-residency requirements.

### Can I run a model privately?

Yes. A private endpoint gives you a reserved GPU or node with the model of your choice, including fine-tuned and Arabic-adapted variants, your own key and limits, and the region you need.

## Related

[BroadNet AI platform Enterprise AI agents, knowledge engine and voice AI.](https://www.broadnet.me/platforms/broadnet-ai) [DialerBee Multilingual AI contact center platform.](https://www.broadnet.me/dialerbee) [The BroadNet Group Every brand in the group, with links.](https://www.broadnet.me/about#group)

## Start with $5 of free credit

Create an account, add a key and send your first Arabic voice note in minutes. Accounts, keys and billing live at inference.broadnet.ai.

[Create an account (opens in a new tab)](https://inference.broadnet.ai/) [Talk to BroadNet](https://www.broadnet.me/contact)

## Related guides & answers

Go deeper on the concepts behind this product.

- [What is conversational commerce Read](https://www.broadnet.me/answers/what-is-conversational-commerce)
- [What is conversational AI in messaging Read](https://www.broadnet.me/answers/what-is-conversational-ai)
- [What is a chatbot for business messaging Read](https://www.broadnet.me/answers/what-is-a-chatbot-for-messaging)
- [How does conversational AI for voice work Read](https://www.broadnet.me/answers/conversational-ai-voice)