FluidInference/speaker-diarization-ov

Pyannote and wespeaker models converted for Speaker diarization and identification for OpenVINO

Load Audio File

import librosa
import matplotlib.pyplot as plt
import librosa.display
import IPython.display as ipd

sample_file = "tutorials_assets_sample.wav"
audio, sr = librosa.load(sample_file)
waveform = torch.from_numpy(audio[0:160000]).unsqueeze(0).unsqueeze(0)
plt.figure(figsize=(14, 5))
librosa.display.waveshow(audio, sr=sr)

ipd.Audio(sample_file)

Loading the pyannote model

core = ov.Core()
model = core.read_model("pyannote-segmentation.xml")
compiled_model = core.compile_model(model, "NPU")  # or "NPU" if supported
input_name = compiled_model.input(0)
output_name = compiled_model.output(0)

results = compiled_model({input_name: waveform})
output = results[output_name]
output.sum(axis=1)

Loading the embedding model

core = ov.Core()
embedding_openvino_model = core.read_model("pyannote-wespeaker.xml")
embedding_openvino_model.reshape((1, 100, 80))
compiled_model = core.compile_model(embedding_openvino_model, "NPU")  # or "NPU" if supported
input_name = compiled_model.input(0)
output_name = compiled_model.output(0)

results = compiled_model({input_name: torch.zeros((1, 100, 80))})
output = results[output_name]
output.sum(axis=1)