Spaces:
Sleeping
Sleeping
Added thread lock to prevent overlapping executions
Browse files- src/app.py +26 -20
src/app.py
CHANGED
@@ -1,6 +1,7 @@
|
|
1 |
import os
|
2 |
|
3 |
from pathlib import Path
|
|
|
4 |
|
5 |
import pandas as pd
|
6 |
|
@@ -30,6 +31,7 @@ processor = AutoProcessor.from_pretrained("facebook/seamless-m4t-v2-large")
|
|
30 |
model = SeamlessM4Tv2Model.from_pretrained("facebook/seamless-m4t-v2-large")
|
31 |
default_sampling_rate = 16_000
|
32 |
|
|
|
33 |
|
34 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
35 |
|
@@ -51,26 +53,30 @@ def translate_audio(
|
|
51 |
:yield: the tuple containing the sampling rate and the audio array
|
52 |
:rtype: tuple[int, np.ndarray]
|
53 |
"""
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
|
|
|
|
|
|
|
|
74 |
|
75 |
|
76 |
# Supported target languages for speech
|
|
|
1 |
import os
|
2 |
|
3 |
from pathlib import Path
|
4 |
+
from threading import Lock
|
5 |
|
6 |
import pandas as pd
|
7 |
|
|
|
31 |
model = SeamlessM4Tv2Model.from_pretrained("facebook/seamless-m4t-v2-large")
|
32 |
default_sampling_rate = 16_000
|
33 |
|
34 |
+
translate_lock = Lock()
|
35 |
|
36 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
37 |
|
|
|
53 |
:yield: the tuple containing the sampling rate and the audio array
|
54 |
:rtype: tuple[int, np.ndarray]
|
55 |
"""
|
56 |
+
with translate_lock:
|
57 |
+
orig_freq, np_array = audio
|
58 |
+
waveform = torch.from_numpy(np_array)
|
59 |
+
waveform = waveform.to(torch.float32)
|
60 |
+
waveform = waveform / 32768.0 # normalize int16 to [-1, 1]
|
61 |
+
|
62 |
+
audio = torchaudio.functional.resample(
|
63 |
+
waveform, orig_freq=orig_freq, new_freq=default_sampling_rate
|
64 |
+
) # must be a 16 kHz waveform array
|
65 |
+
|
66 |
+
audio_inputs = processor(
|
67 |
+
audios=audio,
|
68 |
+
return_tensors="pt",
|
69 |
+
sampling_rate=default_sampling_rate,
|
70 |
+
)
|
71 |
+
|
72 |
+
audio_array_from_audio = (
|
73 |
+
model.generate(**audio_inputs, tgt_lang=tgt_language)[0]
|
74 |
+
.cpu()
|
75 |
+
.numpy()
|
76 |
+
.squeeze()
|
77 |
+
)
|
78 |
+
|
79 |
+
yield (default_sampling_rate, audio_array_from_audio)
|
80 |
|
81 |
|
82 |
# Supported target languages for speech
|