Removed thread lock, changed concurrency limit and added time limit

src/app.py

@@ -1,14 +1,11 @@
 import os
 
 from pathlib import Path
-from threading import Lock
 
 import pandas as pd
-
 import torchaudio
 import torch
 import numpy as np
-
 import gradio as gr
 
 from dotenv import load_dotenv
@@ -18,7 +15,6 @@ from fastrtc import (
     WebRTC,
     ReplyOnPause,
 )
-
 from transformers import AutoProcessor, SeamlessM4Tv2Model
 
 
@@ -31,7 +27,6 @@ processor = AutoProcessor.from_pretrained("facebook/seamless-m4t-v2-large")
 model = SeamlessM4Tv2Model.from_pretrained("facebook/seamless-m4t-v2-large")
 default_sampling_rate = 16_000
 
-translate_lock = Lock()
 
 HF_TOKEN = os.getenv("HF_TOKEN")
 
@@ -53,30 +48,26 @@ def translate_audio(
     :yield: the tuple containing the sampling rate and the audio array
     :rtype: tuple[int, np.ndarray]
     """
-    with translate_lock:
-        orig_freq, np_array = audio
-        waveform = torch.from_numpy(np_array)
-        waveform = waveform.to(torch.float32)
-        waveform = waveform / 32768.0  # normalize int16 to [-1, 1]
-
-        audio = torchaudio.functional.resample(
-            waveform, orig_freq=orig_freq, new_freq=default_sampling_rate
-        )  # must be a 16 kHz waveform array
-
-        audio_inputs = processor(
-            audios=audio,
-            return_tensors="pt",
-            sampling_rate=default_sampling_rate,
-        )
+    orig_freq, np_array = audio
+    waveform = torch.from_numpy(np_array)
+    waveform = waveform.to(torch.float32)
+    waveform = waveform / 32768.0  # normalize int16 to [-1, 1]
+
+    audio = torchaudio.functional.resample(
+        waveform, orig_freq=orig_freq, new_freq=default_sampling_rate
+    )  # must be a 16 kHz waveform array
+
+    audio_inputs = processor(
+        audios=audio,
+        return_tensors="pt",
+        sampling_rate=default_sampling_rate,
+    )
 
-        audio_array_from_audio = (
-            model.generate(**audio_inputs, tgt_lang=tgt_language)[0]
-            .cpu()
-            .numpy()
-            .squeeze()
-        )
+    audio_array_from_audio = (
+        model.generate(**audio_inputs, tgt_lang=tgt_language)[0].cpu().numpy().squeeze()
+    )
 
-        yield (default_sampling_rate, audio_array_from_audio)
+    yield (default_sampling_rate, audio_array_from_audio)
 
 
 # Supported target languages for speech
@@ -215,7 +206,8 @@ with gr.Blocks(
                     ReplyOnPause(translate_audio),
                     inputs=[audio, target_lang],
                     outputs=[audio],
-                    concurrency_limit=1,  # Important: Set to 1 to prevent overlapping executions
+                    concurrency_limit=5,
+                    time_limit=60,
                 )
 
     # Sticky footer (will stay at bottom on all screen sizes)