[Performance Regression] funasr 1.3.1 offline(vad+asr) latency is ~10x slower than 1.3.0 on RTX 4080/4090

[Caxson]

[Possible Regression] funasr==1.3.1 is ~10x slower than 1.3.0 in offline VAD+ASR (RTX 4080/4090)

Summary

After upgrading from funasr==1.3.0 to funasr==1.3.1, we see a large latency increase in offline mode (VAD + ASR).

Under the same code, same audio, same models, and same hardware:

• 1.3.0: ASR latency is typically around 70–80 ms for clips under 10s
• 1.3.1: ASR latency is typically around ~800 ms for the same clips

This looks like a performance regression on our side.

Environment

• GPU: RTX 4080 / RTX 4090
• OS: [please fill]
• Python: [please fill]
• PyTorch: [please fill]
• CUDA + Driver: [please fill]
• FunASR versions compared: 1.3.0 vs 1.3.1

Pipeline

• Mode: offline
• VAD + ASR (+ punctuation)
• Same model IDs and same model revisions across both versions
• Timestamp output disabled for speed (output_timestamp=False)

Reproduction Steps

1. Keep the exact same server code and same short audio set (<10s).
2. Run with funasr==1.3.0 and collect latency logs.
3. Upgrade only FunASR to funasr==1.3.1.
4. Re-run with identical parameters and compare inference_ms / total_ms.

Observed

• 1.3.0: low latency (about 70–80 ms for short clips)
• 1.3.1: much higher latency (about 800 ms on the same clips)

Expected

1.3.1 should have similar latency to 1.3.0 in this offline VAD+ASR setup, or there should be a clear migration note / config change needed to avoid this slowdown.

What we already checked

• Same hardware and runtime setup
• Same code path and same input audio
• Same model IDs/revisions
• Same inference options (including timestamp-disabled path)
• Downgrading to 1.3.0 restores expected latency

Question

Is there any known performance change in 1.3.1 for offline VAD+ASR?
If needed, I can provide:

• minimal reproducible script
• full timing logs
• exact dependency versions (pip freeze)