Titouan commited on
Commit
e1fee45
1 Parent(s): b02bcdd
Files changed (4) hide show
  1. asr.ckpt +3 -0
  2. hyperparams.yaml +67 -0
  3. tokenizer.ckpt +3 -0
  4. wav2vec2.ckpt +3 -0
asr.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d859bb4fa9e63d6645c8f4693437cf09896133a95f036d53d48d1a61ea443d7e
3
+ size 8526821
hyperparams.yaml ADDED
@@ -0,0 +1,67 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # ################################
2
+ # Model: wav2vec2 + DNN + CTC
3
+ # Augmentation: SpecAugment
4
+ # Authors: Sung-Lin Yeh 2021, Titouan Parcollet 2022
5
+ # ################################
6
+
7
+ # URL for the biggest Fairseq english wav2vec2 model.
8
+ wav2vec2_hub: facebook/wav2vec2-large-960h-lv60-self
9
+ sample_rate: 16000
10
+
11
+ # Model parameters
12
+ activation: !name:torch.nn.LeakyReLU
13
+ dnn_layers: 2
14
+ dnn_neurons: 1024
15
+ freeze_wav2vec: True
16
+
17
+ # Outputs
18
+ output_neurons: 31 # BPE size, index(blank/eos/bos) = 0
19
+
20
+ # Decoding parameters
21
+ blank_index: 0
22
+ bos_index: 1
23
+ eos_index: 2
24
+
25
+ enc: !new:speechbrain.lobes.models.VanillaNN.VanillaNN
26
+ input_shape: [null, null, 1024]
27
+ activation: !ref <activation>
28
+ dnn_blocks: !ref <dnn_layers>
29
+ dnn_neurons: !ref <dnn_neurons>
30
+
31
+ wav2vec2: !new:speechbrain.lobes.models.huggingface_wav2vec.HuggingFaceWav2Vec2
32
+ source: !ref <wav2vec2_hub>
33
+ output_norm: True
34
+ freeze: True
35
+ save_path: model_checkpoints
36
+
37
+ ctc_lin: !new:speechbrain.nnet.linear.Linear
38
+ input_size: !ref <dnn_neurons>
39
+ n_neurons: !ref <output_neurons>
40
+
41
+ log_softmax: !new:speechbrain.nnet.activations.Softmax
42
+ apply_log: True
43
+
44
+ ctc_cost: !name:speechbrain.nnet.losses.ctc_loss
45
+ blank_index: !ref <blank_index>
46
+
47
+ asr_model: !new:torch.nn.ModuleList
48
+ - [!ref <enc>, !ref <ctc_lin>]
49
+
50
+ tokenizer: !new:speechbrain.dataio.encoder.CTCTextEncoder
51
+
52
+ encoder: !new:speechbrain.nnet.containers.LengthsCapableSequential
53
+ wav2vec2: !ref <wav2vec2>
54
+ enc: !ref <enc>
55
+ ctc_lin: !ref <ctc_lin>
56
+
57
+ decoding_function: !name:speechbrain.decoders.ctc_greedy_decode
58
+ blank_id: !ref <blank_index>
59
+
60
+ modules:
61
+ encoder: !ref <encoder>
62
+
63
+ pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
64
+ loadables:
65
+ wav2vec2: !ref <wav2vec2>
66
+ asr: !ref <asr_model>
67
+ tokenizer: !ref <tokenizer>
tokenizer.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:28129425176b22cdc33e21f8b2edfdffc74b41ff441e54ec6daa8d5d739ab320
3
+ size 426
wav2vec2.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c4c9a695eee6c61eb6c0452d693db18e682642042525a88ca80430673a90abc8
3
+ size 1261920693