import argparse import json import os from os.path import join as p_join from tqdm import tqdm from time import time import hdbscan import numpy as np import matplotlib.pyplot as plt import seaborn as sns from sklearn.manifold import TSNE import pandas as pd from datasets import load_dataset from model_meta_voice import MetaVoiceEmbedding from model_pyannote_embedding import PyannoteEmbedding from model_clap import CLAPEmbedding, CLAPGeneralEmbedding from model_speaker_embedding import ( W2VBERTEmbedding, Wav2VecEmbedding, XLSR300MEmbedding, XLSR1BEmbedding, XLSR2BEmbedding, HuBERTBaseEmbedding, HuBERTLargeEmbedding, HuBERTXLEmbedding ) def get_embedding(model_class, model_name: str, dataset_name: str, data_split: str): dataset = load_dataset(dataset_name, split=data_split) file_path = p_join("experiment_cache", "embeddings", f"{model_name}.{os.path.basename(dataset_name)}.json") os.makedirs(os.path.dirname(file_path), exist_ok=True) if os.path.exists(file_path): return model = model_class() embeddings = [] for i in tqdm(dataset, total=len(dataset)): start = time() v = model.get_speaker_embedding(i["audio"]["array"], i["audio"]["sampling_rate"]) tmp = { "model": model_name, "embedding": v.tolist(), "sampling_rate": i["audio"]["sampling_rate"], "process_time": time() - start, "dataset_name": os.path.basename(dataset_name) } tmp.update({k: v for k, v in i.items() if k != "audio"}) embeddings.append(tmp) with open(file_path, "w") as f: f.write("\n".join([json.dumps(i) for i in embeddings])) def cluster_embedding(model_name, dataset_name, label_name: str): file_path_embedding = p_join("experiment_cache", "embeddings", f"{model_name}.{os.path.basename(dataset_name)}.json") file_path_cluster = p_join("experiment_cache", "cluster", f"{model_name}.{os.path.basename(dataset_name)}.{label_name}.csv") if not os.path.exists(file_path_cluster): print('CLUSTERING') os.makedirs(os.path.dirname(file_path_cluster), exist_ok=True) assert os.path.exists(file_path_embedding) with open(file_path_embedding) as f: data = [json.loads(i) for i in f.readlines()] clusterer = hdbscan.HDBSCAN() embeddings = [i["embedding"] for i in data] keys = [i for i in range(len(data))] clusterer.fit(np.stack(embeddings)) # data x dimension print(f'{clusterer.labels_.max()} clusters found from {len(data)} data points') print(f"generating report for {label_name}") label = [i[label_name] for i in data] cluster_info = [ {"id": k, "cluster": c, f"label.{label_name}": l} for c, k, l in zip(clusterer.labels_, keys, label) if c != -1 ] cluster_df = pd.DataFrame(cluster_info) cluster_df.to_csv(file_path_cluster, index=False) file_path_tsne = p_join("experiment_cache", "tsne", f"{model_name}.{os.path.basename(dataset_name)}.{label_name}.npy") if not os.path.exists(file_path_tsne): os.makedirs(os.path.dirname(file_path_tsne), exist_ok=True) print('DIMENSION REDUCTION') assert os.path.exists(file_path_embedding) with open(file_path_embedding) as f: data = np.stack([json.loads(i)['embedding'] for i in f.readlines()]) # data x dimension print(f'Dimension reduction: {data.shape}') embedding_2d = TSNE(n_components=2, random_state=0).fit_transform(data) np.save(file_path_tsne, embedding_2d) embedding_2d = np.load(file_path_tsne) print('PLOT') figure_path = p_join("experiment_cache", "figure", f"2d.latent_space.{model_name}.{os.path.basename(dataset_name)}.{label_name}.png") os.makedirs(os.path.dirname(figure_path), exist_ok=True) with open(file_path_embedding) as f: label = np.stack([json.loads(i)[label_name] for i in f.readlines()]) # data x dimension label_type = sorted(list(set(label))) label2id = {v: n for n, v in enumerate(label_type)} plt.figure() scatter = plt.scatter( embedding_2d[:, 0], embedding_2d[:, 1], s=8, c=[label2id[i] for i in label], cmap=sns.color_palette('Spectral', len(label_type), as_cmap=True) ) plt.gca().set_aspect('equal', 'datalim') plt.legend(handles=scatter.legend_elements(num=len(label_type))[0], labels=label_type, bbox_to_anchor=(1.04, 1), borderaxespad=0, loc='upper left', ncol=3 if len(label2id) > 12 else 1) plt.savefig(figure_path, bbox_inches='tight', dpi=600) def main(dataset_name, data_split, label_name): get_embedding(MetaVoiceEmbedding, "meta_voice_se", dataset_name, data_split) cluster_embedding("meta_voice_se", dataset_name, label_name) get_embedding(PyannoteEmbedding, "pyannote_se", dataset_name, data_split) cluster_embedding("pyannote_se", dataset_name, label_name) get_embedding(CLAPEmbedding, "clap_se", dataset_name, data_split) cluster_embedding("clap_se", dataset_name, label_name) get_embedding(CLAPGeneralEmbedding, "clap_general_se", dataset_name, data_split) cluster_embedding("clap_general_se", dataset_name, label_name) get_embedding(HuBERTBaseEmbedding, "hubert_base_se", dataset_name, data_split) cluster_embedding("hubert_base_se", dataset_name, label_name) get_embedding(HuBERTXLEmbedding, "hubert_xl_se", dataset_name, data_split) cluster_embedding("hubert_xl_se", dataset_name, label_name) get_embedding(HuBERTLargeEmbedding, "hubert_large_se", dataset_name, data_split) cluster_embedding("hubert_large_se", dataset_name, label_name) get_embedding(Wav2VecEmbedding, "wav2vec_se", dataset_name, data_split) cluster_embedding("wav2vec_se", dataset_name, label_name) get_embedding(W2VBERTEmbedding, "w2v_bert_se", dataset_name, data_split) cluster_embedding("w2v_bert_se", dataset_name, label_name) get_embedding(XLSR300MEmbedding, "xlsr_300m_se", dataset_name, data_split) cluster_embedding("xlsr_300m_se", dataset_name, label_name) get_embedding(XLSR1BEmbedding, "xlsr_1b_se", dataset_name, data_split) cluster_embedding("xlsr_1b_se", dataset_name, label_name) get_embedding(XLSR2BEmbedding, "xlsr_2b_se", dataset_name, data_split) cluster_embedding("xlsr_2b_se", dataset_name, label_name) if __name__ == '__main__': # main("asahi417/voxceleb1-test-split", "test", "speaker_id") # main("ylacombe/expresso", "train", "speaker_id") # main("ylacombe/expresso", "train", "style") # main("asahi417/j-tube-speech", "test", "speaker_id") main("asahi417/jvnv-emotional-speech-corpus", "test", "style") main("asahi417/jvnv-emotional-speech-corpus", "test", "speaker_id")