Meeting minutes look like one task, but they are two: turning speech into text, and turning text into decisions and action items. Each step fails in its own way. A dialect word mis-transcribed becomes a wrong action item; a model that summarises well can still invent an owner. This post builds the pipeline for Arabic meetings and shows where to put the checks.
What does the pipeline look like?
recording (.m4a, .mp4, .webm)
→ ffmpeg: 16 kHz mono, split at silences
→ speech to text, language=ar, one request per chunk
→ join the transcript, clean obvious noise
→ split the transcript into sections
→ LLM: summary, decisions, action items, open questions (JSON) per section
→ LLM: merge the sections into one set of minutes
→ human review → send
The rest of this post goes step by step.
Step 1: how do you transcribe a long Arabic meeting?
A one-hour meeting recorded on a laptop is usually far bigger than an upload limit. BroadNet Inference takes WAV or MP3 up to 26,214,400 bytes per file, so convert to 16 kHz mono and split at silences first. Transcribe long audio over 25 MB has the full script, including parallel requests and joining the text in order.
Three settings matter for Arabic meetings:
- Send
language="ar"on every chunk. BroadNet’s docs warn that Arabic can otherwise come back transliterated into Latin letters. - Use the Arabic model for Arabic-led meetings:
broadnet/cohere-transcribe-arabic. For meetings mostly in another of its languages,broadnet/cohere-transcribecosts $0.002 a minute. - Keep the raw transcript. You need it for review and for the quotes in step 4.
Mixed Arabic and English is normal in Gulf business meetings. Test a recording with your usual mix: English product names and numbers are where errors show up. Arabic dialect speech recognition covers what to listen for.
Can you get speaker names?
Not from a plain-text transcript. BroadNet’s transcription response is JSON with one text field: no speakers and no timestamps. (For caption timing from chunk offsets, see Arabic subtitles from a text-only API.) You have three options:
| Situation | What to do |
|---|---|
| Each person recorded on a separate track or channel | Transcribe each track on its own; the track tells you the speaker |
| Two-channel call recording | Split the channels with ffmpeg, as in the call center pipeline |
| One mixed track, and you need speakers | Use a provider that returns diarization; the buyer’s guide compares them |
With one mixed track, the LLM can still find owners when someone says a name: “Khalid will send the quote by Thursday.” What it must not do is guess an owner from context. The prompt below forbids that.
Step 2: why split the transcript before summarising?
Two reasons. First, context: an hour of talk can be too long for a small model. BroadNet Gemma 4 VL 26B has a 32,768-token context, which a long Arabic meeting can exceed. Count yours with the Arabic token counter; Arabic usually needs more tokens per word than English, as tokens per word in Arabic vs English shows.
Second, accuracy. Research on long inputs found that model performance “significantly degrades when models must access relevant information in the middle of long contexts” (Liu et al.). A decision made in minute 35 of 70 is exactly that. Summarising sections of 10 to 15 minutes each, then merging, keeps every part near the start of a prompt.
Split on the same silence points you used for the audio, so a section never starts mid-sentence.
Step 3: which model should write the minutes?
| Model | Context | Notes for minutes | Input / output per 1M tokens |
|---|---|---|---|
| Qwen3.8 27B | 262,144 | Tested by BroadNet in Arabic: fluent writing, Gulf and Egyptian dialects understood; structured output listed | $0.175 / $2.17 |
| GLM 5.3 | 1,048,576 | Structured output listed; text only | $0.78 / $3.46 |
| DeepSeek V4.1 Flash | 1,048,576 | Also on a flat plan at $60 a month per session, one request at a time | $0.175 / $0.52 |
Start with Qwen3.8 27B, because it is the model BroadNet tested in Arabic. Test it against one other model on three real meetings, and have a native speaker score them. Choosing an LLM for Arabic has a scoring sheet.
All three models reason before they answer, and the reasoning counts toward max_tokens. Minutes are long, so allow several thousand tokens, or the visible answer can come back empty. Why that happens.
Step 4: what prompt gives reliable minutes?
Ask for JSON, and make the model quote the transcript for every item. The quote is what lets a reviewer check an action item in seconds instead of rereading the whole transcript.
import json, os
from openai import OpenAI
client = OpenAI(base_url="https://inference.broadnet.ai/v1", api_key=os.environ["BROADNET_API_KEY"])
SYSTEM = """You write meeting minutes from an Arabic meeting transcript section.
Return JSON only, with these keys:
"summary": 2-4 sentences in Modern Standard Arabic,
"decisions": [{"text": ..., "quote": ...}],
"actions": [{"task": ..., "owner": ..., "due": ..., "quote": ...}],
"open_questions": [{"text": ..., "quote": ...}]
Rules:
- "quote" is the exact words from the transcript that support the item.
- "owner" is a name only if the transcript says who will do it. Otherwise null.
- "due" only if a date or day is said. Keep it as said, e.g. "الخميس". Otherwise null.
- Keep names, numbers and amounts exactly as in the transcript. Never guess."""
def minutes_for(section):
r = client.chat.completions.create(
model="broadnet/qwen3.8-27b",
max_tokens=4000,
messages=[{"role": "system", "content": SYSTEM},
{"role": "user", "content": section}])
return json.loads(r.choices[0].message.content)
Then validate in code. Parse the JSON, check the keys, and check that every quote really appears in the section. An item whose quote is not in the transcript is the model’s invention: drop it or flag it. If the JSON does not parse, retry once with the error message. JSON mode vs json_schema shows a validation loop; BroadNet’s docs list structured output for GLM 5.3 and Qwen3.8 27B without naming the request parameter, so test it before relying on it.
For minutes in English, change the summary line to English and add: “Keep each quote in the original Arabic.” The Arabic prompt engineering post covers register, numerals and mixed-script text in more depth.
Step 5: how do you merge sections into one set of minutes?
Send all the section results in one final request and ask the model to:
- Write one summary of the whole meeting.
- Merge duplicate decisions and actions. The same task is often mentioned twice.
- Keep each item’s quote.
- Mark an action as changed if a later section changes its owner or date.
The merge input is small: JSON from each section, not the transcript. A merge on any of the three models above stays far inside its context.
What does one meeting cost?
| Step | What you pay for | Price |
|---|---|---|
| Transcription | Audio minutes | $0.0042 a minute, $0.25 an hour |
| Section summaries | Transcript tokens in, JSON out | Qwen3.8 27B: $0.175 in, $2.17 out per 1M |
| Merge | Small JSON in and out | Same rates |
Transcription is usually the fixed part. The LLM part depends on transcript length and on how much the model reasons. Read prompt_tokens and completion_tokens from the first few real meetings, then multiply in the LLM API cost calculator. The speech-to-text cost calculator does the audio side. The starting credit of $5 covers the transcription models, which is about 1,190 minutes of Arabic; Qwen3.8 27B needs a top-up.
What mistakes should you avoid?
- Summarising the whole meeting in one prompt. Items from the middle go missing.
- Letting the model fill gaps. Owners and dates that were never said are the most common error in AI minutes. Make null the default.
- Translating too early. Transcribe and extract in Arabic, then translate the minutes. Translating the transcript first loses the exact quotes.
- Skipping review. Minutes are a record. A person who attended should read them before they go out.
- Keeping recordings forever. Delete audio once the minutes are approved, unless you must keep it.
When is an API pipeline the wrong choice?
If you need live captions during the meeting, speaker labels from one mixed track, or a ready-made app your team can use today, a dedicated meeting product is the faster route. Several sell Arabic meeting notes; compare them on your own recordings, since accuracy claims on their pages are their own. The API route makes sense when you want minutes inside your own system, in your format, with your own retention rules.
Where BroadNet Inference fits
BroadNet puts both halves on one OpenAI-compatible key: BroadNet Transcribe Arabic at $0.0042 a minute, and Arabic-tested Qwen3.8 27B for the minutes. Its privacy statement says it keeps no prompts, audio, answers or transcripts after a request, which matters for internal meetings. It does not return speaker labels or timestamps, and it has no live transcription, so use another provider if you need those. Start with the Arabic speech-to-text guide, then add the summary step above.
