Rajaram1996 marcmaxmeister commited on
Commit
81e08bf
1 Parent(s): 7c15f78

Adding an example of using pretrained model to predict emotion in local audio file (#1)

Browse files

- Adding an example of using pretrained model to predict emotion in local audio file (f02201ba227deefa8868a872db7ca70556ae44ef)


Co-authored-by: Marc Maxmeister <[email protected]>

Files changed (1) hide show
  1. README.md +53 -1
README.md CHANGED
@@ -6,4 +6,56 @@ tags:
6
  - audio
7
  - HUBert
8
  ---
9
- A place to hold the model for easier inference.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6
  - audio
7
  - HUBert
8
  ---
9
+
10
+
11
+ Working example of using pretrained model to predict emotion in local audio file
12
+
13
+ ```
14
+
15
+ def predict_emotion_hubert(audio_file):
16
+ """ inspired by an example from https://github.com/m3hrdadfi/soxan """
17
+ from audio_models import HubertForSpeechClassification
18
+ from transformers import Wav2Vec2FeatureExtractor, AutoConfig
19
+ import torch.nn.functional as F
20
+ import torch
21
+ import numpy as np
22
+ from pydub import AudioSegment
23
+
24
+ model = HubertForSpeechClassification.from_pretrained("Rajaram1996/Hubert_emotion") # Downloading: 362M
25
+ feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("facebook/hubert-base-ls960")
26
+ sampling_rate=16000 # defined by the model; must convert mp3 to this rate.
27
+ config = AutoConfig.from_pretrained("Rajaram1996/Hubert_emotion")
28
+
29
+ def speech_file_to_array(path, sampling_rate):
30
+ # using torchaudio...
31
+ # speech_array, _sampling_rate = torchaudio.load(path)
32
+ # resampler = torchaudio.transforms.Resample(_sampling_rate, sampling_rate)
33
+ # speech = resampler(speech_array).squeeze().numpy()
34
+ sound = AudioSegment.from_file(path)
35
+ sound = sound.set_frame_rate(sampling_rate)
36
+ sound_array = np.array(sound.get_array_of_samples())
37
+ return sound_array
38
+
39
+ sound_array = speech_file_to_array(audio_file, sampling_rate)
40
+ inputs = feature_extractor(sound_array, sampling_rate=sampling_rate, return_tensors="pt", padding=True)
41
+ inputs = {key: inputs[key].to("cpu").float() for key in inputs}
42
+
43
+ with torch.no_grad():
44
+ logits = model(**inputs).logits
45
+
46
+ scores = F.softmax(logits, dim=1).detach().cpu().numpy()[0]
47
+ outputs = [{
48
+ "emo": config.id2label[i],
49
+ "score": round(score * 100, 1)}
50
+ for i, score in enumerate(scores)
51
+ ]
52
+ return [row for row in sorted(outputs, key=lambda x:x["score"], reverse=True) if row['score'] != '0.0%'][:2]
53
+ ```
54
+
55
+ ```
56
+
57
+ result = predict_emotion_hubert("male-crying.mp3")
58
+ >>> result
59
+ [{'emo': 'male_sad', 'score': 91.0}, {'emo': 'male_fear', 'score': 4.8}]
60
+ ```
61
+