Spaces:
Paused
Paused
| import tensorflow.keras as keras | |
| import numpy as np | |
| import librosa | |
| import random | |
| #import tensorflow_addons | |
| SAMPLE_RATE = 22050 | |
| class _Emotion_spotting_service(): | |
| model = None | |
| #instance = None | |
| mapping = [' amazement', ' solemnity', ' tenderness', | |
| ' nostalgia', ' calmness', ' power', | |
| ' joyfulness', ' tension',' sadness'] | |
| def __init__(self,model_path): | |
| self.model = keras.models.load_model(model_path) | |
| def predict(self,file_path): | |
| log_spectrogram = self.preprocess(file_path) | |
| X = np.array(log_spectrogram).astype("float32") | |
| X = np.expand_dims(X, axis=0) | |
| # Do predictions | |
| num_predictions = self.model.predict(X) | |
| prediction = np.argmax(num_predictions) | |
| predicted_keyword = self.mapping[prediction] | |
| return predicted_keyword | |
| # Split audio into 10 second excerpts | |
| # Attain log spectrogram with following parameters | |
| # sample rate = 22050, n_fft = 2048, hop_length = 512 | |
| # output, 1024*431 | |
| def preprocess(self,file_path): | |
| signal, sr = librosa.load(file_path,sr=SAMPLE_RATE) | |
| signal_normalized = librosa.util.normalize(signal) | |
| len_to_check = 10 * 22050 | |
| # If audio is less than 10 seconds, we pad it with zeroes | |
| # If audio is more than 10 seconds, we split into segments and randomly choose one | |
| if len(signal_normalized) < len_to_check: | |
| num_zeros = len_to_check - len(signal_normalized) | |
| signal_normalized = signal_normalized + [0] * num_zeros | |
| elif len(signal_normalized) > len_to_check: | |
| num_segments = len(signal_normalized)//len_to_check | |
| segments = [] | |
| for i in range(num_segments): | |
| start = i * len_to_check | |
| end = start + len_to_check | |
| if len(signal[start:end]) != len_to_check: | |
| continue | |
| else: | |
| segments.append(signal[start:end]) | |
| signal_normalized = random.choice(segments) | |
| stft = librosa.stft(signal_normalized, n_fft=2048,hop_length=512)[:-1] | |
| spectrogram = np.abs(stft) | |
| log_spectrogram = librosa.amplitude_to_db(spectrogram) | |
| return log_spectrogram | |
| # def Emotion_spotting_service(): | |
| # if _Emotion_spotting_service.instance == None: | |
| # _Emotion_spotting_service.instance = _Emotion_spotting_service() | |
| # _Emotion_spotting_service.model = keras.models.load_model("ERM.h5") | |
| # return _Emotion_spotting_service.instance | |
| # if __name__ == "__main__": | |
| # emotion_service = _Emotion_spotting_service("emotion_model.h5") | |
| # predicted_word = emotion_service.predict("10.mp3") | |
| # print(predicted_word) |