DmitrySobolevsky
commited on
Commit
•
b8500cb
1
Parent(s):
7b257e3
Adding handler
Browse files- handler.py +35 -0
handler.py
ADDED
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Dict
|
2 |
+
from transformers.pipelines.audio_utils import ffmpeg_read
|
3 |
+
import whisper
|
4 |
+
import torch
|
5 |
+
|
6 |
+
SAMPLE_RATE = 16000
|
7 |
+
|
8 |
+
|
9 |
+
class EndpointHandler():
|
10 |
+
def __init__(self, path=""):
|
11 |
+
# load the model
|
12 |
+
self.model = whisper.load_model("large")
|
13 |
+
|
14 |
+
def __call__(self, data: Dict[str, bytes]) -> Dict[str, str]:
|
15 |
+
"""
|
16 |
+
Args:
|
17 |
+
data (:obj:):
|
18 |
+
includes the deserialized audio file as bytes
|
19 |
+
Return:
|
20 |
+
A :obj:`dict`:. base64 encoded image
|
21 |
+
"""
|
22 |
+
# process input
|
23 |
+
inputs = data.pop("inputs", data)
|
24 |
+
audio_nparray = ffmpeg_read(inputs, SAMPLE_RATE)
|
25 |
+
audio_tensor= torch.from_numpy(audio_nparray)
|
26 |
+
|
27 |
+
# run inference pipeline
|
28 |
+
result = self.model.transcribe(audio_nparray)
|
29 |
+
|
30 |
+
# postprocess the prediction
|
31 |
+
return {
|
32 |
+
"segments": result["segments"],
|
33 |
+
"language_code": result["language"],
|
34 |
+
"text": result["text"],
|
35 |
+
}
|