# we have to evaluate emotion & cer per sentence -> not audinterface sliding window import os import audresample import torch import matplotlib.pyplot as plt import soundfile import json import audb from transformers import AutoModelForAudioClassification from transformers.models.wav2vec2.modeling_wav2vec2 import Wav2Vec2PreTrainedModel import types from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline import pandas as pd import json import numpy as np from pathlib import Path import transformers import torch import audmodel import audiofile import jiwer # https://arxiv.org/pdf/2407.12229 # https://arxiv.org/pdf/2312.05187 # https://arxiv.org/abs/2407.05407 # https://arxiv.org/pdf/2408.06577 # https://arxiv.org/pdf/2309.07405 import msinference import os from random import shuffle config = transformers.Wav2Vec2Config() #finetuning_task='spef2feat_reg') config.dev = torch.device('cuda:0') config.dev2 = torch.device('cuda:0') LABELS = ['arousal', 'dominance', 'valence', 'Angry', 'Sad', 'Happy', 'Surprise', 'Fear', 'Disgust', 'Contempt', 'Neutral' ] config = transformers.Wav2Vec2Config() #finetuning_task='spef2feat_reg') config.dev = torch.device('cuda:0') config.dev2 = torch.device('cuda:0') # https://arxiv.org/pdf/2407.12229 # https://arxiv.org/pdf/2312.05187 # https://arxiv.org/abs/2407.05407 # https://arxiv.org/pdf/2408.06577 # https://arxiv.org/pdf/2309.07405 def _infer(self, x): '''x: (batch, audio-samples-16KHz)''' x = (x + self.config.mean) / self.config.std # plus x = self.ssl_model(x, attention_mask=None).last_hidden_state # pool h = self.pool_model.sap_linear(x).tanh() w = torch.matmul(h, self.pool_model.attention) w = w.softmax(1) mu = (x * w).sum(1) x = torch.cat( [ mu, ((x * x * w).sum(1) - mu * mu).clamp(min=1e-7).sqrt() ], 1) return self.ser_model(x) teacher_cat = AutoModelForAudioClassification.from_pretrained( '3loi/SER-Odyssey-Baseline-WavLM-Categorical-Attributes', trust_remote_code=True # fun definitions see 3loi/SER-.. repo ).to(config.dev2).eval() teacher_cat.forward = types.MethodType(_infer, teacher_cat) # ===================[:]===================== Dawn def _prenorm(x, attention_mask=None): '''mean/var''' if attention_mask is not None: N = attention_mask.sum(1, keepdim=True) # here attn msk is unprocessed just the original input x -= x.sum(1, keepdim=True) / N var = (x * x).sum(1, keepdim=True) / N else: x -= x.mean(1, keepdim=True) # mean is an onnx operator reducemean saves some ops compared to casting integer N to float and the div var = (x * x).mean(1, keepdim=True) return x / torch.sqrt(var + 1e-7) from torch import nn from transformers.models.wav2vec2.modeling_wav2vec2 import Wav2Vec2PreTrainedModel, Wav2Vec2Model class RegressionHead(nn.Module): r"""Classification head.""" def __init__(self, config): super().__init__() self.dense = nn.Linear(config.hidden_size, config.hidden_size) self.dropout = nn.Dropout(config.final_dropout) self.out_proj = nn.Linear(config.hidden_size, config.num_labels) def forward(self, features, **kwargs): x = features x = self.dropout(x) x = self.dense(x) x = torch.tanh(x) x = self.dropout(x) x = self.out_proj(x) return x class Dawn(Wav2Vec2PreTrainedModel): r"""Speech emotion classifier.""" def __init__(self, config): super().__init__(config) self.config = config self.wav2vec2 = Wav2Vec2Model(config) self.classifier = RegressionHead(config) self.init_weights() def forward( self, input_values, attention_mask=None, ): x = _prenorm(input_values, attention_mask=attention_mask) outputs = self.wav2vec2(x, attention_mask=attention_mask) hidden_states = outputs[0] hidden_states = torch.mean(hidden_states, dim=1) logits = self.classifier(hidden_states) return logits # return {'hidden_states': hidden_states, # 'logits': logits} dawn = Dawn.from_pretrained('audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim').to(config.dev).eval() # ======================================= torch_dtype = torch.float16 #if torch.cuda.is_available() else torch.float32 model_id = "openai/whisper-large-v3" model = AutoModelForSpeechSeq2Seq.from_pretrained( model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True ).to(config.dev) processor = AutoProcessor.from_pretrained(model_id) _pipe = pipeline( "automatic-speech-recognition", model=model, tokenizer=processor.tokenizer, feature_extractor=processor.feature_extractor, max_new_tokens=128, chunk_length_s=30, batch_size=16, return_timestamps=True, torch_dtype=torch_dtype, device=config.dev, ) def process_function(x, sampling_rate, idx): # x = x[None , :] ASaHSuFDCN # {0: 'Angry', 1: 'Sad', 2: 'Happy', 3: 'Surprise', # 4: 'Fear', 5: 'Disgust', 6: 'Contempt', 7: 'Neutral'} #tensor([[0.0015, 0.3651, 0.0593, 0.0315, 0.0600, 0.0125, 0.0319, 0.4382]]) logits_cat = teacher_cat(torch.from_numpy(x).to(config.dev)).softmax(1) logits_adv = dawn(torch.from_numpy(x).to(config.dev)) out = torch.cat([logits_adv, logits_cat], 1).cpu().detach().numpy() # print(out.shape) return out[0, :] def load_speech(split=None): DB = [ # [dataset, version, table, has_timdeltas_or_is_full_wavfile] # ['crema-d', '1.1.1', 'emotion.voice.test', False], #['librispeech', '3.1.0', 'test-clean', False], ['emodb', '1.2.0', 'emotion.categories.train.gold_standard', False], # ['entertain-playtestcloud', '1.1.0', 'emotion.categories.train.gold_standard', True], # ['erik', '2.2.0', 'emotion.categories.train.gold_standard', True], # ['meld', '1.3.1', 'emotion.categories.train.gold_standard', False], # ['msppodcast', '5.0.0', 'emotion.categories.train.gold_standard', False], # tandalone bucket because it has gt labels? # ['myai', '1.0.1', 'emotion.categories.train.gold_standard', False], # ['casia', None, 'emotion.categories.gold_standard', False], # ['switchboard-1', None, 'sentiment', True], # ['swiss-parliament', None, 'segments', True], # ['argentinian-parliament', None, 'segments', True], # ['austrian-parliament', None, 'segments', True], # #'german', --> bundestag # ['brazilian-parliament', None, 'segments', True], # ['mexican-parliament', None, 'segments', True], # ['portuguese-parliament', None, 'segments', True], # ['spanish-parliament', None, 'segments', True], # ['chinese-vocal-emotions-liu-pell', None, 'emotion.categories.desired', False], # peoples-speech slow # ['peoples-speech', None, 'train-initial', False] ] output_list = [] for database_name, ver, table, has_timedeltas in DB: a = audb.load(database_name, sampling_rate=16000, format='wav', mixdown=True, version=ver, cache_root='/cache/audb/') a = a[table].get() if has_timedeltas: print(f'{has_timedeltas=}') # a = a.reset_index()[['file', 'start', 'end']] # output_list += [[*t] for t # in zip(a.file.values, a.start.dt.total_seconds().values, a.end.dt.total_seconds().values)] else: output_list += [f for f in a.index] # use file (no timedeltas) return output_list natural_wav_paths = load_speech() with open('harvard.json', 'r') as f: harvard_individual_sentences = json.load(f)['sentences'] synthetic_wav_paths = ['./enslow/' + i for i in os.listdir('./enslow/')] synthetic_wav_paths_4x = ['./style_vector_v2/' + i for i in os.listdir('./style_vector_v2/')] synthetic_wav_paths_foreign = ['./mimic3_foreign/' + i for i in os.listdir('./mimic3_foreign/') if 'en_U' not in i] synthetic_wav_paths_foreign_4x = ['./mimic3_foreign_4x/' + i for i in os.listdir('./mimic3_foreign_4x/') if 'en_U' not in i] # very short segments # filter very short styles synthetic_wav_paths_foreign = [i for i in synthetic_wav_paths_foreign if audiofile.duration(i) > 2] synthetic_wav_paths_foreign_4x = [i for i in synthetic_wav_paths_foreign_4x if audiofile.duration(i) > 2] synthetic_wav_paths = [i for i in synthetic_wav_paths if audiofile.duration(i) > 2] synthetic_wav_pathsn_4x = [i for i in synthetic_wav_paths_4x if audiofile.duration(i) > 2] shuffle(synthetic_wav_paths_foreign_4x) shuffle(synthetic_wav_paths_foreign) shuffle(synthetic_wav_paths) shuffle(synthetic_wav_paths_4x) print(len(synthetic_wav_paths_foreign_4x), len(synthetic_wav_paths_foreign), len(synthetic_wav_paths), len(synthetic_wav_paths_4x)) # 134 204 134 204 for audio_prompt in ['english', 'english_4x', 'human', 'foreign', 'foreign_4x']: # each of these creates a separate pkl - so outer for # data = np.zeros((770, len(LABELS)*2 + 2)) # 768 x LABELS-prompt & LABELS-stts2 & cer-prompt & cer-stts2 # OUT_FILE = f'{audio_prompt}_analytic.pkl' if not os.path.isfile(OUT_FILE): ix = 0 for list_of_10 in harvard_individual_sentences[:10004]: # long_sentence = ' '.join(list_of_10['sentences']) # harvard.append(long_sentence.replace('.', ' ')) for text in list_of_10['sentences']: if audio_prompt == 'english': _p = synthetic_wav_paths[ix % len(synthetic_wav_paths)] # 134 style_vec = msinference.compute_style(_p) elif audio_prompt == 'english_4x': _p = synthetic_wav_paths_4x[ix % len(synthetic_wav_paths_4x)] # 134] style_vec = msinference.compute_style(_p) elif audio_prompt == 'human': _p = natural_wav_paths[ix % len(natural_wav_paths)] # ? style_vec = msinference.compute_style(_p) elif audio_prompt == 'foreign': _p = synthetic_wav_paths_foreign[ix % len(synthetic_wav_paths_foreign)] # 204 some short styles are discarded ~ 1180 style_vec = msinference.compute_style(_p) elif audio_prompt == 'foreign_4x': _p = synthetic_wav_paths_foreign_4x[ix % len(synthetic_wav_paths_foreign_4x)] # 174 style_vec = msinference.compute_style(_p) else: print('unknonw list of style vector') x = msinference.inference(text, style_vec, alpha=0.3, beta=0.7, diffusion_steps=7, embedding_scale=1) x = audresample.resample(x, 24000, 16000) _st, fsr = audiofile.read(_p) _st = audresample.resample(_st, fsr, 16000) print(_st.shape, x.shape) emotion_of_prompt = process_function(_st, 16000, None) emotion_of_out = process_function(x, 16000, None) data[ix, :11] = emotion_of_prompt data[ix, 11:22] = emotion_of_out # 2 last columns is cer-prompt cer-styletts2 transcription_prompt = _pipe(_st[0]) transcription_styletts2 = _pipe(x[0]) # allow singleton for EMO process func # print(len(emotion_of_prompt + emotion_of_out), ix, text) print(transcription_prompt, transcription_styletts2) data[ix, 22] = jiwer.cer('Sweet dreams are made of this. I travel the world and the seven seas.', transcription_prompt['text']) data[ix, 23] = jiwer.cer(text, transcription_styletts2['text']) print(data[ix, :]) ix += 1 df = pd.DataFrame(data, columns=['prompt-' + i for i in LABELS] + ['styletts2-' + i for i in LABELS] + ['cer-prompt', 'cer-styletts2']) df.to_pickle(OUT_FILE) else: df = pd.read_pickle(OUT_FILE) print('\nALREADY EXISTS\n{df}') # From the pickle we should also run cer and whisper on every prompt