My BosonAI's hackathon project ranked 2nd place in benchmarking stream - Benchmarking the scalable Audio Higgs Model and Qwen3.2
My BosonAI’s hackathon project ranked 2nd place in benchmarking stream: Benchmarking the scalable Audio Higgs Model + Qwen3.2 
Project Overview
Problem
When customers call support, they often get forced into a rigid menu (“press 1/2/3…”), even though:
- they don’t know which category their issue belongs to,
- routing takes time,
- scaling human agents is expensive,
- requests can span multiple domains.
Our idea
Use audio-first routing:
- synthesize/ingest caller request (TTS caller),
- understand request → decide department + intent,
- dispatch to the right expert model,
- respond as an agent via TTS.
System Diagram
System pipeline: audio request → understanding/router → expert(s) → TTS agent response.
Audio Examples (Caller vs Agent)
Below are paired audio clips (caller request → agent response).
Tip: keep filenames short + lowercase.
Example 1 — Simple QA
Caller: “What is two plus two?”
Agent: short phrase answer
Example 7 — Time question
Caller: “How many minutes are in an hour?”
Agent: short phrase answer
Case Study: Airline Customer Calling Service
Example 0
“Please help me make a new reservation for Toronto.”
Example 35
“Online check-in isn’t working for me.”
Example 56
“Set up UM service for my child.”
Routing taxonomy (skeleton)
We route into:
- Department (BOOKING / CHANGES / REFUNDS / …)
- Intention (NEW_BOOKING / SEAT_SELECTION / …)
Click to expand routing labels
- **BOOKING**: NEW_BOOKING, SEAT_SELECTION, … - **CHANGES**: FLIGHT_CHANGE, NAME_CORRECTION, … - **REFUNDS**: CANCEL_ITINERARY, VOUCHER_QUESTION, … - **BAGGAGE**: ADD_EXTRA_BAGS, LOST_BAGS, … - **FLIGHT_STATUS**: STATUS_REQUEST, DELAY_REASON, … - **CHECK_IN**: ONLINE_CHECKIN, PASSPORT_ISSUE, … - **LOYALTY**: MISSING_MILES, MERGE_ACCOUNTS, … - **PAYMENT**: PAYMENT_FAILED, REQUEST_RECEIPT, … - **SPECIAL_ASSISTANCE**: WHEELCHAIR, SPECIAL_MEAL, … - **GENERAL**: SPEAK_TO_AGENT, OTHER, …Benchmarking: Concurrency
Latency p50/p95 by stage vs concurrency"
Key metrics (fill in)
| Setting | p50 Total Latency (ms) | p95 Total Latency (ms) | Notes |
|---|---|---|---|
| C=1 | TBD | TBD | baseline |
| C=2 | TBD | TBD | |
| C=4 | TBD | TBD | |
| C=8 | TBD | TBD |
Benchmarking: Small dataset
Confusion matrices on SMALL DATASET (department / intent).
Summary numbers (example)
- Dataset: 64 examples
- Department accuracy: 87.5%
- Intent accuracy: 87.5%
Confusion matrices on LARGE DATASET (department / intent).
Reproducibility
Environment
- TTS caller: HiggAudio v2
- Router: HiggAudioUnderstand + Qwen3.2
- TTS agent: HiggAudio v2
How to run (skeleton)
# 1) install
pip install -r requirements.txt
# 2) run demo
python demo.py --config configs/demo.yaml