recorded_audio buffer grows without bounds causing progressive latency

[Lounes78]

Hello,
When running the model (tested with both CLI and Streamlit clients), the audio works well for ~3 seconds, then degrades into laggy, distorted output. Latency keeps increasing.

Latency rises from ~70–110 ms to over 200 ms and keeps growing:

Data sent ... Data received ... Received in 73.86 ms
Data sent ... Data received ... Received in 112.37 ms
...
Data sent ... Data received ... Received in 144.70 ms
...
Data sent ... Data received ... Received in 209.72 ms
...
Data sent ... Data received ... Received in 213.94 ms

Eventually reaching 250-270ms+

Setup

• GPU: Quadro RTX 6000 (24GB)
• CUDA: 12.7
• Python: 3.12

Server Logs Sample

recorded audio buffer grows

Recorded audio shape torch.Size([2, 138000]), audio tensor shape torch.Size([1, 2, 2000])
Recorded audio shape torch.Size([2, 140000]), audio tensor shape torch.Size([1, 2, 2000])
Recorded audio shape torch.Size([2, 142000]), audio tensor shape torch.Size([1, 2, 2000])
...
Recorded audio shape torch.Size([2, 200000]), audio tensor shape torch.Size([1, 2, 2000])
...
Recorded audio shape torch.Size([2, 300000]), audio tensor shape torch.Size([1, 2, 2000])
...
Recorded audio shape torch.Size([2, 386000]), audio tensor shape torch.Size([1, 2, 2000])

The recorded audio buffer grows from 138k to 386k+ samples

Has anyone encountered this issue and found a solution ?