speech-test commited on
Commit
97ffc8c
1 Parent(s): c1297a0

Upload model

Browse files
Files changed (4) hide show
  1. README.md +96 -0
  2. config.json +107 -0
  3. preprocessor_config.json +9 -0
  4. pytorch_model.bin +3 -0
README.md ADDED
@@ -0,0 +1,96 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: en
3
+ datasets:
4
+ - superb
5
+ tags:
6
+ - speech
7
+ - audio
8
+ - hubert
9
+ - audio-classification
10
+ license: apache-2.0
11
+ ---
12
+
13
+ # Hubert-Large for Keyword Spotting
14
+
15
+ ## Model description
16
+
17
+ This is a ported version of
18
+ [S3PRL's Hubert for the SUPERB Keyword Spotting task](https://github.com/s3prl/s3prl/tree/master/s3prl/downstream/speech_commands).
19
+
20
+ The base model is [hubert-large-ll60k](https://huggingface.co/facebook/hubert-large-ll60k), which is pretrained on 16kHz
21
+ sampled speech audio. When using the model make sure that your speech input is also sampled at 16Khz.
22
+
23
+ For more information refer to [SUPERB: Speech processing Universal PERformance Benchmark](https://arxiv.org/abs/2105.01051)
24
+
25
+ ## Task and dataset description
26
+
27
+ Keyword Spotting (KS) detects preregistered keywords by classifying utterances into a predefined set of
28
+ words. The task is usually performed on-device for the fast response time. Thus, accuracy, model size, and
29
+ inference time are all crucial. SUPERB uses the widely used
30
+ [Speech Commands dataset v1.0](https://www.tensorflow.org/datasets/catalog/speech_commands) for the task.
31
+ The dataset consists of ten classes of keywords, a class for silence, and an unknown class to include the
32
+ false positive.
33
+
34
+ For the original model's training and evaluation instructions refer to the
35
+ [S3PRL downstream task README](https://github.com/s3prl/s3prl/tree/master/s3prl/downstream#ks-keyword-spotting).
36
+
37
+
38
+ ## Usage examples
39
+
40
+ You can use the model via the Audio Classification pipeline:
41
+ ```python
42
+ from datasets import load_dataset
43
+ from transformers import pipeline
44
+
45
+ dataset = load_dataset("anton-l/superb_demo", "ks", split="test")
46
+
47
+ classifier = pipeline("audio-classification", model="superb/hubert-large-superb-ks")
48
+ labels = classifier(dataset[0]["file"], top_k=5)
49
+ ```
50
+
51
+ Or use the model directly:
52
+ ```python
53
+ import torch
54
+ from datasets import load_dataset
55
+ from transformers import HubertForSequenceClassification, Wav2Vec2FeatureExtractor
56
+ from torchaudio.sox_effects import apply_effects_file
57
+
58
+ effects = [["channels", "1"], ["rate", "16000"], ["gain", "-3.0"]]
59
+ def map_to_array(example):
60
+ speech, _ = apply_effects_file(example["file"], effects)
61
+ example["speech"] = speech.squeeze(0).numpy()
62
+ return example
63
+
64
+ # load a demo dataset and read audio files
65
+ dataset = load_dataset("anton-l/superb_demo", "ks", split="test")
66
+ dataset = dataset.map(map_to_array)
67
+
68
+ model = HubertForSequenceClassification.from_pretrained("superb/hubert-large-superb-ks")
69
+ feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("superb/hubert-large-superb-ks")
70
+
71
+ # compute attention masks and normalize the waveform if needed
72
+ inputs = feature_extractor(dataset[:4]["speech"], sampling_rate=16000, padding=True, return_tensors="pt")
73
+
74
+ logits = model(**inputs).logits
75
+ predicted_ids = torch.argmax(logits, dim=-1)
76
+ labels = [model.config.id2label[_id] for _id in predicted_ids.tolist()]
77
+ ```
78
+
79
+ ## Eval results
80
+
81
+ The evaluation metric is accuracy.
82
+
83
+ | | **s3prl** | **transformers** |
84
+ |--------|-----------|------------------|
85
+ |**test**| `0.9529` | `N/A` |
86
+
87
+ ### BibTeX entry and citation info
88
+
89
+ ```bibtex
90
+ @article{yang2021superb,
91
+ title={SUPERB: Speech processing Universal PERformance Benchmark},
92
+ author={Yang, Shu-wen and Chi, Po-Han and Chuang, Yung-Sung and Lai, Cheng-I Jeff and Lakhotia, Kushal and Lin, Yist Y and Liu, Andy T and Shi, Jiatong and Chang, Xuankai and Lin, Guan-Ting and others},
93
+ journal={arXiv preprint arXiv:2105.01051},
94
+ year={2021}
95
+ }
96
+ ```
config.json ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/hubert-large-ll60k",
3
+ "activation_dropout": 0.0,
4
+ "apply_spec_augment": true,
5
+ "architectures": [
6
+ "HubertForSequenceClassification"
7
+ ],
8
+ "attention_dropout": 0.1,
9
+ "bos_token_id": 1,
10
+ "classifier_proj_size": 256,
11
+ "conv_bias": true,
12
+ "conv_dim": [
13
+ 512,
14
+ 512,
15
+ 512,
16
+ 512,
17
+ 512,
18
+ 512,
19
+ 512
20
+ ],
21
+ "conv_kernel": [
22
+ 10,
23
+ 3,
24
+ 3,
25
+ 3,
26
+ 3,
27
+ 2,
28
+ 2
29
+ ],
30
+ "conv_stride": [
31
+ 5,
32
+ 2,
33
+ 2,
34
+ 2,
35
+ 2,
36
+ 2,
37
+ 2
38
+ ],
39
+ "ctc_loss_reduction": "sum",
40
+ "ctc_zero_infinity": false,
41
+ "do_stable_layer_norm": true,
42
+ "eos_token_id": 2,
43
+ "feat_extract_activation": "gelu",
44
+ "feat_extract_dropout": 0.0,
45
+ "feat_extract_norm": "layer",
46
+ "feat_proj_dropout": 0.1,
47
+ "final_dropout": 0.0,
48
+ "gradient_checkpointing": false,
49
+ "hidden_act": "gelu",
50
+ "hidden_dropout": 0.1,
51
+ "hidden_size": 1024,
52
+ "id2label": {
53
+ "0": "yes",
54
+ "1": "no",
55
+ "2": "up",
56
+ "3": "down",
57
+ "4": "left",
58
+ "5": "right",
59
+ "6": "on",
60
+ "7": "off",
61
+ "8": "stop",
62
+ "9": "go",
63
+ "10": "_unknown_",
64
+ "11": "_silence_"
65
+ },
66
+ "initializer_range": 0.02,
67
+ "intermediate_size": 4096,
68
+ "label2id": {
69
+ "_silence_": 11,
70
+ "_unknown_": 10,
71
+ "down": 3,
72
+ "go": 9,
73
+ "left": 4,
74
+ "no": 1,
75
+ "off": 7,
76
+ "on": 6,
77
+ "right": 5,
78
+ "stop": 8,
79
+ "up": 2,
80
+ "yes": 0
81
+ },
82
+ "layer_norm_eps": 1e-05,
83
+ "layerdrop": 0.1,
84
+ "mask_channel_length": 10,
85
+ "mask_channel_min_space": 1,
86
+ "mask_channel_other": 0.0,
87
+ "mask_channel_prob": 0.0,
88
+ "mask_channel_selection": "static",
89
+ "mask_feature_length": 10,
90
+ "mask_feature_prob": 0.0,
91
+ "mask_time_length": 10,
92
+ "mask_time_min_space": 1,
93
+ "mask_time_other": 0.0,
94
+ "mask_time_prob": 0.075,
95
+ "mask_time_selection": "static",
96
+ "model_type": "hubert",
97
+ "num_attention_heads": 16,
98
+ "num_conv_pos_embedding_groups": 16,
99
+ "num_conv_pos_embeddings": 128,
100
+ "num_feat_extract_layers": 7,
101
+ "num_hidden_layers": 24,
102
+ "pad_token_id": 0,
103
+ "torch_dtype": "float32",
104
+ "transformers_version": "4.11.0.dev0",
105
+ "use_weighted_layer_sum": true,
106
+ "vocab_size": 32
107
+ }
preprocessor_config.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "do_normalize": false,
3
+ "feature_extractor_type": "Wav2Vec2FeatureExtractor",
4
+ "feature_size": 1,
5
+ "padding_side": "right",
6
+ "padding_value": 0,
7
+ "return_attention_mask": true,
8
+ "sampling_rate": 16000
9
+ }
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c504591a9c6e209ce7d45d047a46d68e8de0ca498f187bf2d4cfc03b923df036
3
+ size 1262979115