patrickvonplaten commited on
Commit
97f6728
1 Parent(s): d07fd10

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +47 -82
README.md CHANGED
@@ -1,96 +1,61 @@
1
- # Wav2Vec2 Acoustic Model fine-tuned on LibriSpeech
 
 
 
 
 
2
 
3
- Original model can be found under https://github.com/pytorch/fairseq/tree/master/examples/wav2vec#wav2vec-20.
 
4
 
5
- Paper: https://arxiv.org/abs/2006.11477
6
 
7
- ## Usage
8
 
9
- Make sure you are working on [this branch](https://github.com/huggingface/transformers/tree/add_wav2vec) (which will be merged to master soon hopefully) of transformers:
10
 
11
- ```bash
12
- $ git checkout add_wav2vec
13
- ```
14
 
15
- In the following, we'll show a simple example of how the model can be used for automatic speech recognition.
16
 
17
- First, let's load the model
18
 
19
- ```python
20
- from transformers import AutoModelForMaskedLM
21
-
22
- model = AutoModelForMaskedLM.from_pretrained("patrickvonplaten/wav2vec2-large-960h")
23
-
24
- ```
25
-
26
- Next, let's load a dummy librispeech dataset
27
-
28
- ```python
29
- from datasets import load_dataset
30
- import soundfile as sf
31
-
32
- libri_speech_dummy = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
33
-
34
- def map_to_array(batch):
35
- speech_array, _ = sf.read(batch["file"])
36
- batch["speech"] = speech_array
37
- return batch
38
-
39
- libri_speech_dummy = libri_speech_dummy.map(map_to_array, remove_columns=["file"])
40
-
41
- # check out dataset
42
- print(libri_speech_dummy)
43
 
44
- input_speech_16kHz = libri_speech_dummy[2]["speech"]
45
- expected_trans = libri_speech_dummy[2]["text"]
46
- ```
47
 
48
- Cool, now we can run an inference pass to retrieve the logits:
49
 
50
- ```python
51
- import torch
52
- logits = model(torch.tensor(input_speech_16kHz)[None, :])
53
-
54
- # use highest probability logits
55
- pred_ids = torch.argmax(logits[0], axis=-1)
56
- ```
57
-
58
- Finally, let's decode the prediction.
59
- Let's create a simple CTC-Decoder:
60
-
61
- ```python
62
- import numpy as np
63
- from itertools import groupby
64
 
65
- class Decoder:
66
- def __init__(self, json_dict):
67
- self.dict = json_dict
68
- self.look_up = np.asarray(list(self.dict.keys()))
69
-
70
- def decode(self, ids):
71
- converted_tokens = self.look_up[ids]
72
- fused_tokens = [tok[0] for tok in groupby(converted_tokens)]
73
- output = ' '.join(''.join(''.join(fused_tokens).split("<s>")).split("|"))
74
- return output
75
- ```
76
-
77
- and instantiate with the corresponding dict.
78
 
79
  ```python
80
- # hard-coded json dict taken from: https://dl.fbaipublicfiles.com/fairseq/wav2vec/dict.ltr.txt
81
- json_dict = {"<s>": 0, "<pad>": 1, "</s>": 2, "<unk>": 3, "|": 4, "E": 5, "T": 6, "A": 7, "O": 8, "N": 9, "I": 10, "H": 11, "S": 12, "R": 13, "D": 14, "L": 15, "U": 16, "M": 17, "W": 18, "C": 19, "F": 20, "G": 21, "Y": 22, "P": 23, "B": 24, "V": 25, "K": 26, "'": 27, "X": 28, "J": 29, "Q": 30, "Z": 31}
82
-
83
- decoder = Decoder(json_dict=json_dict)
84
- ```
85
-
86
- and decode the result
87
-
88
- ```python
89
- pred_trans = decoder.decode(pred_ids)
90
-
91
- print("Prediction:\n", pred_trans)
92
- print("\n" + 50 * "=" + "\n")
93
- print("Correct result:\n", expected_trans)
94
- ```
95
-
96
- 🎉
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: en
3
+ datasets:
4
+ - librispeech_asr
5
+ tags:
6
+ - speech
7
 
8
+ license: apache-2.0
9
+ ---
10
 
11
+ # Wav2Vec2-Base-960h
12
 
13
+ [Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/)
14
 
15
+ The large model pretrained and fine-tuned on 960 hours of Librispeech.
16
 
17
+ [Paper](https://arxiv.org/abs/2006.11477)
 
 
18
 
19
+ Authors: Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli
20
 
21
+ **Abstract**
22
 
23
+ We show for the first time that learning powerful representations from speech audio alone followed by fine-tuning on transcribed speech can outperform the best semi-supervised methods while being conceptually simpler. wav2vec 2.0 masks the speech input in the latent space and solves a contrastive task defined over a quantization of the latent representations which are jointly learned. Experiments using all labeled data of Librispeech achieve 1.8/3.3 WER on the clean/other test sets. When lowering the amount of labeled data to one hour, wav2vec 2.0 outperforms the previous state of the art on the 100 hour subset while using 100 times less labeled data. Using just ten minutes of labeled data and pre-training on 53k hours of unlabeled data still achieves 4.8/8.2 WER. This demonstrates the feasibility of speech recognition with limited amounts of labeled data.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
24
 
25
+ The original model can be found under https://github.com/pytorch/fairseq/tree/master/examples/wav2vec#wav2vec-20.
 
 
26
 
 
27
 
28
+ # Usage
 
 
 
 
 
 
 
 
 
 
 
 
 
29
 
30
+ To transcribe audio files the model can be used as a standalone acoustic model as follows:
 
 
 
 
 
 
 
 
 
 
 
 
31
 
32
  ```python
33
+ from transformers import Wav2Vec2Tokenizer, Wav2Vec2Model
34
+ from datasets import load_dataset
35
+ import soundfile as sf
36
+ import torch
37
+
38
+ # load model and tokenizer
39
+ tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-large-960h")
40
+ model = Wav2Vec2ForMaskedLM.from_pretrained("facebook/wav2vec2-large-960h")
41
+
42
+ # define function to read in sound file
43
+ def map_to_array(batch):
44
+ speech, _ = sf.read(batch["file"])
45
+ batch["speech"] = speech
46
+ return batch
47
+
48
+ # load dummy dataset and read soundfiles
49
+ ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
50
+ ds = ds.map(map_to_array)
51
+
52
+ # tokenize
53
+ input_values = tokenizer(ds["speech"][:2], return_tensors="pt", padding="longest").input_values # Batch size 1
54
+
55
+ # retrieve logits
56
+ logits = model(input_values).logits
57
+
58
+ # take argmax and decode
59
+ predicted_ids = torch.argmax(logits, dim=-1)
60
+ transcription = tokenizer.batch_decode(predicted_ids)
61
+ ```