Audio_CNN1D / example_usage.py

Upload folder using huggingface_hub

b5c4dba verified over 1 year ago

2.09 kB

import tensorflow as tf
import numpy as np
import json
import librosa
import os

def load_model(model_path):
    "Load the emotion recognition model."
    return tf.keras.models.load_model(model_path)

def predict_emotion(model, audio_path, preprocessor_config):
    "Predict emotion from an audio file."
    # Load audio file
    waveform, sr = librosa.load(
        audio_path, 
        sr=preprocessor_config["sample_rate"],
        duration=preprocessor_config["duration"],
        offset=preprocessor_config["offset"]
    )
    
    # Ensure consistent length
    target_length = int(preprocessor_config["sample_rate"] * preprocessor_config["duration"])
    if len(waveform) < target_length:
        waveform = np.pad(waveform, (0, target_length - len(waveform)))
    if len(waveform) > target_length:
        waveform = waveform[:target_length]
    
    # Extract features
    mel_spec = librosa.feature.melspectrogram(
        y=waveform,
        sr=preprocessor_config["sample_rate"],
        n_fft=preprocessor_config["frame_length"],
        hop_length=preprocessor_config["hop_length"],
        n_mels=128
    )
    
    # Convert to log scale
    log_mel = np.log(np.maximum(mel_spec, 1e-10))
    
    # Add batch and channel dimensions
    features = np.expand_dims(np.expand_dims(log_mel, axis=0), axis=-1)
    
    # Make prediction
    prediction = model.predict(features)[0]
    
    # Get emotion label
    emotion_idx = np.argmax(prediction)
    emotion = preprocessor_config["emotions"][str(emotion_idx)]
    
    return emotion, prediction[emotion_idx]

# Example usage
if __name__ == "__main__":
    # Load model
    model = load_model("emotion_recognition_model.keras")
    
    # Load preprocessor_config
    with open("preprocessing.json", "r") as f:
        preprocessor_config = json.load(f)
    
    # Path to your audio file
    audio_path = "path/to/your/audio.wav"
    
    # Predict emotion
    emotion, confidence = predict_emotion(model, audio_path, preprocessor_config)
    
    print(f"Predicted emotion: {emotion} with confidence {confidence:.2f}")


	import tensorflow as tf
	import numpy as np
	import json
	import librosa
	import os

	def load_model(model_path):
	"Load the emotion recognition model."
	return tf.keras.models.load_model(model_path)

	def predict_emotion(model, audio_path, preprocessor_config):
	"Predict emotion from an audio file."
	# Load audio file
	waveform, sr = librosa.load(
	audio_path,
	sr=preprocessor_config["sample_rate"],
	duration=preprocessor_config["duration"],
	offset=preprocessor_config["offset"]
	)

	# Ensure consistent length
	target_length = int(preprocessor_config["sample_rate"] * preprocessor_config["duration"])
	if len(waveform) < target_length:
	waveform = np.pad(waveform, (0, target_length - len(waveform)))
	if len(waveform) > target_length:
	waveform = waveform[:target_length]

	# Extract features
	mel_spec = librosa.feature.melspectrogram(
	y=waveform,
	sr=preprocessor_config["sample_rate"],
	n_fft=preprocessor_config["frame_length"],
	hop_length=preprocessor_config["hop_length"],
	n_mels=128
	)

	# Convert to log scale
	log_mel = np.log(np.maximum(mel_spec, 1e-10))

	# Add batch and channel dimensions
	features = np.expand_dims(np.expand_dims(log_mel, axis=0), axis=-1)

	# Make prediction
	prediction = model.predict(features)[0]

	# Get emotion label
	emotion_idx = np.argmax(prediction)
	emotion = preprocessor_config["emotions"][str(emotion_idx)]

	return emotion, prediction[emotion_idx]

	# Example usage
	if __name__ == "__main__":
	# Load model
	model = load_model("emotion_recognition_model.keras")

	# Load preprocessor_config
	with open("preprocessing.json", "r") as f:
	preprocessor_config = json.load(f)

	# Path to your audio file
	audio_path = "path/to/your/audio.wav"

	# Predict emotion
	emotion, confidence = predict_emotion(model, audio_path, preprocessor_config)

	print(f"Predicted emotion: {emotion} with confidence {confidence:.2f}")