Recordings are uploaded to the host as WAV files, timed from upload to finished transcript and speaker diarization; audio comes from the public FLEURS (CC BY 4.0) and AISHELL-1 (Apache 2.0) datasets. Core model is Alibaba DAMO Academy SenseVoice-Small (multilingual: zh/en/ja/ko/yue) + CAM++ speaker diarization.
1. Real-Time Single-Badge Interaction (30 s Dialogue)
| Scenario & Sample | Audio Duration | Total Processing Time | Speaker Separation | Hardware Option | Site Experience |
|---|---|---|---|---|---|
| Two-speaker Chinese service (FLEURS) | 26.1 s | 1.21 s | 2 speakers / 5 turns | reComputer J40 Series (J4012) | Near-instant: text appears right after dialogue ends |
| Two-speaker long Chinese (FLEURS) | 30.5 s | 1.17 s | 2 speakers / 4 turns | reComputer J40 Series (J4012) | 175x real-time speed, imperceptible latency |
| Two-speaker Chinese (AISHELL-1) | 16.5 s | 1.25 s | 2 speakers / 3 turns | reComputer J30 Series (J3011) | Mainstream cost-effective option, same experience as J40 |
| Japanese counter audio (FLEURS) | 34.0 s | 1.71 s | 2 Japanese turns | reComputer J40 Series (J4012) | Rapid multilingual handling for customer reviews |
| Two-speaker Chinese (FLEURS) | 30.5 s | 4.61 s | 2 speakers / 4 turns | reComputer RK3576 Series (8 GB) | Under 5 s, ideal for non-urgent consultation logs |
2. Shift-End Batch Upload Concurrency (10 Badges Uploading 10 min Total Audio)
| Hardware Option | Recommended Concurrency | 10 Badges Total Time | Max Queue Waiting Latency | Reliability & Scheduling Mechanism | Recommended Deployment Site |
|---|---|---|---|---|---|
| reComputer J40 Series<br>(Orin NX 16GB) | 10 ~ 50 badges | ~3.5 s | < 3.5 s (imperceptible) | Sub-second queue drain: Single-context CUDA Graph lock protects against OOM and 429s | Supermarkets, chain stores, civic service halls |
| reComputer J30 Series<br>(Orin Nano 8GB) | 5 ~ 20 badges | ~4.1 s | < 4.1 s (imperceptible) | Same architecture, GPU throughput 29.5 QPS, only 486 MB VRAM | Standard retail stores with 8~15 staff |
| reComputer RK3588 Series<br>(RK3588 8GB) | 5 ~ 15 badges | ~14.5 s | ~5.0 s (3 parallel pipelines) | Hardware multi-core concurrency: 3 independent NPU cores run 3 parallel workers | Mid-size workstations, multi-counter hubs |
| reComputer RK3576 Series<br>(RK3576 8GB) | 1 ~ 3 badges | ~46 s | ~46 s (sequential) | Lightweight single-worker: Peak RSS only 442 MB (W4A16), sequential processing | Private clinics, counseling rooms, small counters |
For multilingual latency, speaker diarization benchmarks and RK3576 memory metrics, see Engineering Wiki · Smart Badge Voice Transcription.