Upload handler.py with huggingface_hub
Browse files- handler.py +47 -0
handler.py
ADDED
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Dict
|
2 |
+
from pyannote.audio import Pipeline
|
3 |
+
from transformers.pipelines.audio_utils import ffmpeg_read
|
4 |
+
import torch
|
5 |
+
|
6 |
+
SAMPLE_RATE = 16000
|
7 |
+
|
8 |
+
|
9 |
+
|
10 |
+
class EndpointHandler():
|
11 |
+
def __init__(self, path=""):
|
12 |
+
# load the model
|
13 |
+
self.pipeline = Pipeline.from_pretrained("philschmid/pyannote-speaker-diarization-endpoint")
|
14 |
+
|
15 |
+
|
16 |
+
def __call__(self, data: Dict[str, bytes]) -> Dict[str, str]:
|
17 |
+
"""
|
18 |
+
Args:
|
19 |
+
data (:obj:):
|
20 |
+
includes the deserialized audio file as bytes
|
21 |
+
Return:
|
22 |
+
A :obj:`dict`:. base64 encoded image
|
23 |
+
"""
|
24 |
+
# process input
|
25 |
+
inputs = data.pop("inputs", data)
|
26 |
+
parameters = data.pop("parameters", None) # min_speakers=2, max_speakers=5
|
27 |
+
|
28 |
+
|
29 |
+
# prepare pynannote input
|
30 |
+
audio_nparray = ffmpeg_read(inputs, SAMPLE_RATE)
|
31 |
+
audio_tensor= torch.from_numpy(audio_nparray).unsqueeze(0)
|
32 |
+
pyannote_input = {"waveform": audio_tensor, "sample_rate": SAMPLE_RATE}
|
33 |
+
|
34 |
+
# apply pretrained pipeline
|
35 |
+
# pass inputs with all kwargs in data
|
36 |
+
if parameters is not None:
|
37 |
+
diarization = self.pipeline(pyannote_input, **parameters)
|
38 |
+
else:
|
39 |
+
diarization = self.pipeline(pyannote_input)
|
40 |
+
|
41 |
+
# postprocess the prediction
|
42 |
+
processed_diarization = [
|
43 |
+
{"label": str(label), "start": str(segment.start), "stop": str(segment.end)}
|
44 |
+
for segment, _, label in diarization.itertracks(yield_label=True)
|
45 |
+
]
|
46 |
+
|
47 |
+
return {"diarization": processed_diarization}
|