In much of the Middle East, customers do not type. They send a 40-second voice note in Gulf or Egyptian Arabic, often with an order number read out in the middle. Agents then listen, rewind, and type it into the helpdesk. Multiply by a few hundred a day and voice notes become the slowest channel you have.
This post builds the business workflow around transcription: getting notes out of WhatsApp and Telegram, turning them into structured tickets, and keeping the data under control. For the conversion and transcription code itself, see our step-by-step guide to transcribing WhatsApp voice notes. We will not repeat it here.
How does a voice note become a support ticket?
- Receive the webhook from WhatsApp or Telegram.
- Download the audio immediately.
- Convert to 16 kHz mono WAV.
- Transcribe with the right language code.
- Extract ticket fields with an LLM.
- Create the ticket, with transcript, and reply in text.
- Delete the audio file.
Steps 1 and 2: how do you get the audio out of WhatsApp?
On the WhatsApp Business Cloud API, an inbound voice note arrives as a webhook with a media ID, not the file. Meta’s media docs set the rules:
- Call
GET /<MEDIA_ID>on the Graph API to get aurl. - Download that URL with
Authorization: Bearer <token>. Without the token the request fails. - “Media URLs expire after 5 minutes.” Request a new one if you are late.
- Media IDs from incoming webhooks are valid for 7 days.
import os, requests
GRAPH = os.environ["GRAPH_BASE"] # e.g. https://graph.facebook.com/<version>
HDR = {"Authorization": f"Bearer {os.environ['WA_TOKEN']}"}
def download_whatsapp(media_id, path):
url = requests.get(f"{GRAPH}/{media_id}", headers=HDR, timeout=10).json()["url"]
with open(path, "wb") as f:
f.write(requests.get(url, headers=HDR, timeout=30).content)
Do this in the webhook handler’s background job, not later in a batch. A queue that runs once an hour will hit expired URLs.
How do you get voice notes out of a Telegram bot?
A Telegram bot receives a voice object with a file_id. The Bot API says bots “can download files of up to 20MB”, via https://api.telegram.org/file/bot<token>/<file_path>, and the link is “valid for at least 1 hour”.
TG = f"https://api.telegram.org/bot{os.environ['TG_TOKEN']}"
def download_telegram(file_id, path):
file_path = requests.get(f"{TG}/getFile", params={"file_id": file_id}, timeout=10).json()["result"]["file_path"]
with open(path, "wb") as f:
f.write(requests.get(f"https://api.telegram.org/file/bot{os.environ['TG_TOKEN']}/{file_path}", timeout=30).content)
Both platforms deliver Ogg/Opus voice notes in practice. The transcription endpoint takes WAV or MP3, so convert with the one-line ffmpeg command from the WhatsApp guide, then transcribe with language="ar" (or the customer’s language).
Step 5: how do you turn a transcript into a ticket?
A raw transcript is still hard to triage. An LLM can pull out the fields your helpdesk needs:
import json
from openai import OpenAI
client = OpenAI(base_url="https://inference.broadnet.ai/v1", api_key=os.environ["BROADNET_API_KEY"])
FIELDS = ("Return JSON with: intent (refund|delivery|billing|technical|other), urgency (low|normal|high), "
"order_number (string or null), language, summary_en (one sentence), needs_human (true/false).")
def to_ticket(transcript):
reply = client.chat.completions.create(
model="broadnet/qwen3.8-27b",
response_format={"type": "json_object"},
messages=[{"role": "system", "content": "You triage customer support voice notes. " + FIELDS},
{"role": "user", "content": transcript}],
).choices[0].message.content
return json.loads(reply)
Qwen3.8 27B is the model BroadNet tested for Arabic, including Gulf and Egyptian dialects. Validate the JSON before you write it to the helpdesk, and fall back to needs_human: true when parsing fails; the structured output post has a validate-and-repair loop. If a note mentions card or ID numbers, redact before this step; see sending sensitive data to an LLM API.
Which edge cases should you handle early?
Real inboxes are messier than a demo. Plan for these from the first week:
- Several notes in a row. Customers often send three short notes instead of one long one. Group notes from the same sender within a couple of minutes into one ticket, and transcribe them in order.
- Very short notes. “Okay” or “thanks” notes do not need a ticket. Skip extraction when the transcript is under a few words.
- Forwarded notes. A forwarded voice note may be someone else speaking. Treat it as an attachment, not as the customer’s own words.
- Mixed language. Gulf customers mix in English product names. If your queue serves several languages, record the language you sent with each transcript so agents know what to expect.
- Large files. Meta’s media docs list a 16 MB audio limit and Telegram bots can download up to 20 MB, both under the 25 MiB transcription limit, but long recordings still take longer. Process them in the background.
What a team gets from it
| Before | After |
|---|---|
| Agent listens to each note, often twice | Agent reads a transcript and a one-line summary |
| Notes cannot be searched | Transcripts are searchable by order number or keyword |
| Triage by whoever is free | Routing by intent and urgency |
| Arabic notes wait for an Arabic speaker | English summary lets any agent pick up simple cases |
What does voice-note transcription cost?
Transcription is the cheap part of the pipeline. At $0.0042 a minute, 10,000 Arabic voice notes of 30 seconds cost $21.00. The 14-language model costs $0.002 a minute. The LLM step is priced per token; a short transcript plus the instruction is typically a few hundred tokens. Put your own volumes into the speech-to-text cost calculator, or compare other providers in the speech-to-text pricing comparison.
How should you handle privacy and retention?
Voice notes are personal data, and they often contain more than the customer meant to share. Three habits help:
- Delete audio after transcription unless a policy requires you to keep it.
- Store transcripts with the ticket, under the helpdesk’s access controls, not in app logs.
- Use a provider that does not keep audio. BroadNet keeps no audio or transcripts after the request; see zero data retention.
Reply in plain text, and keep first replies to new contacts short and free of links. Messaging platforms watch for spam-like patterns from business numbers.
Where BroadNet Inference fits
BroadNet covers the transcription and extraction steps with one OpenAI-compatible key: BroadNet Transcribe Arabic for Arabic notes, BroadNet Transcribe for Arabic plus 13 other languages, and Arabic-tested chat models for the ticket fields. BroadNet reports its Arabic model #2 of 37 (its September 2026 figure; the live board listed it #4 of 39 on 15 November 2026) on the Open Universal Arabic ASR Leaderboard. The $5 starting credit covers about 1,190 minutes of Arabic audio.
BroadNet does not connect to WhatsApp, Telegram or your helpdesk for you; the webhook handling and ticket creation are your code or your automation tool. Calls are a different shape of problem, with two speakers and long recordings: see transcribing call center recordings.

