BroadNet runs a free browser playground for BroadNet Inference. It calls the live API through a shared key, so you can hear the Arabic transcription model on your own voice or test a document before you create an account. BroadNet says request and response content is not retained.
What can you try?
The playground has four modes:
| Mode | What you give it | What comes back |
|---|---|---|
| Text | A question in Arabic or English | A streamed answer from Gemma 4 |
| Vision | One image, such as a menu, form or handwritten note, plus a question | An answer about the image |
| Voice | A recording or upload of up to 60 seconds, Arabic or English | A transcript, then a summary |
| Documents | One image of an invoice, ID card, flyer or other document | The fields as JSON |
Voice accepts WAV up to 4 MB or WebM up to 1.44 MB. That is a playground limit; the API takes WAV or MP3 up to 25 MiB per file.
How do you test Arabic speech to text?
- Open the voice mode.
- Pick العربية for Arabic or English.
- Record up to a minute, or upload a short voice note.
- Run the demo and read the transcript and summary.
Try the audio you will really send: a WhatsApp note recorded outside, a phone call, a dialect you care about. A clean read sentence tells you little; BroadNet’s own numbers show the same model at 3.8 word error rate on clean read speech and 23.4 on dialect speech. See what Arabic WER means.
How do you test document reading?
Open the documents mode, choose the document type, upload one image and run it. Check the numbers against the original: totals, dates, ID numbers and phone numbers are where extraction goes wrong. For the same job over the API, see reading Arabic invoices with a vision model.
How do you move the same test to the API?
The playground is for a first look. For real tests you need your own key:
- Create an account at inference.broadnet.ai. New accounts get $5 of credit with no card.
- Create an API key in your account.
- Send the same request with curl or any OpenAI SDK. For a voice note:
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
Always send language; without it, Arabic can come back in Latin letters. Convert .ogg or .m4a voice notes to WAV first, because the API accepts WAV and MP3 only:
ffmpeg -i voice-note.ogg -ar 16000 -ac 1 voice-note.wav
For chat and images, the curl quickstart has a first request for each endpoint.
What are the limits of a playground test?
- One file at a time. It shows quality, not throughput. Pay-as-you-go keys allow 30 requests a minute and 4 in flight.
- Short clips only. Sixty seconds will not show how a model handles a long call. Test a few full recordings over the API.
- No settings. You cannot set temperature, prompts or response formats. The API exposes the usual OpenAI parameters.
- A shared key. It is a demo, not a service. Do not build on it.
