app.py

from speechbrain.inference.interfaces import foreign_class
from custom_interface import CustomEncoderWav2vec2Classifier
from speechbrain.pretrained import EncoderClassifier
import openvino.properties.hint as hints

# Function in SpeechBrain to load and use custom PyTorch models 
classifier = foreign_class(
   source="speechbrain/emotion-recognition-wav2vec2-IEMOCAP",
   pymodule_file="custom_interface.py",
   classname="CustomEncoderWav2vec2Classifier"
)

# Model checkpoint files
checkpoint = EncoderClassifier.from_hparams(
    source="speechbrain/emotion-recognition-wav2vec2-IEMOCAP",
    savedir="./"  # Directory to save the model
)

# Convert hparams to a dictionary
hparams_dict = vars(checkpoint.hparams)

# inference backend
backend = "openvino"
torch_device = "cpu"

if backend == "openvino" and torch_device == "cpu":

    # OpenVINO inference optimization parameters
    config = {hints.performance_mode: hints.PerformanceMode.THROUGHPUT}
    ov_opts = {"ov_device": "CPU", "config": config}
    instance = CustomEncoderWav2vec2Classifier(modules=checkpoint.mods,
                                           hparams=hparams_dict, model=classifier.mods["wav2vec2"].model,
                                           audio_file_path="speechbrain/emotion-recognition-wav2vec2-IEMOCAP/anger.wav",
                                           backend="openvino",
                                           opts=ov_opts,
                                           torch_device=torch_device,
                                           save_ov_model=False)
elif backend == "openvino" and torch_device == "cuda":
    raise ValueError("OpenVINO backend does not support CUDA devices. \
                            Please use cpu for torch_device.")

if backend == "pytorch":
    torch_opts = {"torch_device": torch_device}
    instance = CustomEncoderWav2vec2Classifier(modules=checkpoint.mods,
                                           hparams=hparams_dict, model=classifier.mods["wav2vec2"].model,
                                           audio_file_path="speechbrain/emotion-recognition-wav2vec2-IEMOCAP/anger.wav",
                                           backend="pytorch", opts=torch_opts, torch_device=torch_device)

# OpenVINO inference
print("=" * 30)
if backend == "openvino":
    print(f"[INFO] Inference Device: {ov_opts['ov_device']}")
    print("=" * 30)
    print("\n[INFO] Performing OpenVINO inference...")
else:
    print(f"[INFO] Inference Device: {torch_opts['torch_device']}")
    print("=" * 30)
    print("\n[INFO] Performing PyTorch inference...")


out_prob, score, index, text_lab = instance.classify_file("speechbrain/emotion-recognition-wav2vec2-IEMOCAP/anger.wav")
print(f"[RESULT] Inference output label: {text_lab[index-1]}")