# requirement packages
!pip install git+https://github.com/huggingface/datasets.git
!pip install git+https://github.com/huggingface/transformers.git
!pip install torchaudio
!pip install librosa

!git clone https://github.com/SeaBenSea/HuBERT-SER.git

Prediction

python

import sys  
sys.path.insert(1, './HuBERT-SER/')
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchaudio
from transformers import AutoConfig, Wav2Vec2FeatureExtractor
from src.models import Wav2Vec2ForSpeechClassification, HubertForSpeechClassification

python

model_name_or_path = "SeaBenSea/hubert-large-turkish-speech-emotion-recognition"

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
config = AutoConfig.from_pretrained(model_name_or_path)
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(model_name_or_path)
sampling_rate = feature_extractor.sampling_rate

model = HubertForSpeechClassification.from_pretrained(model_name_or_path).to(device)

python

def speech_file_to_array_fn(path, sampling_rate):
    speech_array, _sampling_rate = torchaudio.load(path)
    resampler = torchaudio.transforms.Resample(_sampling_rate, sampling_rate)
    speech = resampler(speech_array).squeeze().numpy()
    return speech


def predict(path, sampling_rate):
    speech = speech_file_to_array_fn(path, sampling_rate)
    inputs = feature_extractor(speech, sampling_rate=sampling_rate, return_tensors="pt", padding=True)
    inputs = {key: inputs[key].to(device) for key in inputs}

    with torch.no_grad():
        logits = model(**inputs).logits

    scores = F.softmax(logits, dim=1).detach().cpu().numpy()[0]
    outputs = [{"Emotion": config.id2label[i], "Score": f"{round(score * 100, 3):.1f}%"} for i, score in
               enumerate(scores)]
    return outputs

python

path = "../dataset/TurEV/Angry/1157_kz_acik.wav"
outputs = predict(path, sampling_rate)
outputs

bash

[
  {'Emotion': 'Angry', 'Score': '99.8%'},
  {'Emotion': 'Calm', 'Score': '0.0%'},
  {'Emotion': 'Happy', 'Score': '0.1%'},
  {'Emotion': 'Sad', 'Score': '0.1%'}
]

Evaluation

The following tables summarize the scores obtained by model overall and per each class.

Emotions	precision	recall	f1-score	accuracy
Angry	0.97	0.99	0.98
Calm	0.89	0.95	0.92
Happy	0.98	0.93	0.95
Sad	0.97	0.93	0.95
			Overal	0.95

Questions?

Post a Github issue from HERE.