Files changed (9) hide show
  1. .gitattributes +4 -0
  2. README.md +126 -0
  3. asr.ckpt +3 -0
  4. config.json +3 -0
  5. example.wav +0 -0
  6. hyperparams.yaml +156 -0
  7. lm.ckpt +3 -0
  8. normalizer.ckpt +3 -0
  9. tokenizer.ckpt +3 -0
.gitattributes CHANGED
@@ -30,3 +30,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
30
  *.zip filter=lfs diff=lfs merge=lfs -text
31
  *.zst filter=lfs diff=lfs merge=lfs -text
32
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
30
  *.zip filter=lfs diff=lfs merge=lfs -text
31
  *.zst filter=lfs diff=lfs merge=lfs -text
32
  *tfevents* filter=lfs diff=lfs merge=lfs -text
33
+ normalizer.ckpt filter=lfs diff=lfs merge=lfs -text
34
+ tokenizer.ckpt filter=lfs diff=lfs merge=lfs -text
35
+ asr.ckpt filter=lfs diff=lfs merge=lfs -text
36
+ lm.ckpt filter=lfs diff=lfs merge=lfs -text
README.md CHANGED
@@ -1,3 +1,129 @@
1
  ---
 
 
 
 
 
 
 
 
 
 
2
  license: apache-2.0
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ language:
3
+ - en
4
+ thumbnail: null
5
+ tags:
6
+ - automatic-speech-recognition
7
+ - CTC
8
+ - Attention
9
+ - Transformer
10
+ - pytorch
11
+ - speechbrain
12
  license: apache-2.0
13
+ datasets:
14
+ - switchboard
15
+ metrics:
16
+ - wer
17
+ - cer
18
+
19
  ---
20
+
21
+ <iframe src="https://ghbtns.com/github-btn.html?user=speechbrain&repo=speechbrain&type=star&count=true&size=large&v=2" frameborder="0" scrolling="0" width="170" height="30" title="GitHub"></iframe>
22
+ <br/><br/>
23
+
24
+ # Transformer for Switchboard (with Transformer LM)
25
+
26
+ This repository provides all the necessary tools to perform automatic speech
27
+ recognition from an end-to-end system pretrained on Switchboard (EN) within SpeechBrain.
28
+ For a better experience, we encourage you to learn more about [SpeechBrain](https://speechbrain.github.io).
29
+
30
+ The performance of the model is the following:
31
+
32
+ | Release | Swbd WER | Callhome WER | Eval2000 WER | GPUs |
33
+ |:--------:|:--------:|:------------:|:------------:|:-----------:|
34
+ | 17-09-22 | 9.80 | 17.89 | 13.94 | 1xA100 40GB |
35
+
36
+
37
+ ## Pipeline Description
38
+
39
+ This ASR system is composed of 3 different but linked blocks:
40
+ - Tokenizer (unigram) that transforms words into subword units trained on the Switchboard training transcriptions and the Fisher corpus.
41
+ - Neural language model (Transformer LM) trained on the Switchboard training transcriptions and the Fisher corpus.
42
+ - Acoustic model made of a transformer encoder and a joint decoder with CTC +
43
+ transformer. Hence, the decoding also incorporates the CTC probabilities.
44
+
45
+ The system is trained with recordings sampled at 16kHz (single channel).
46
+ The code will automatically normalize your audio (i.e., resampling + mono channel selection) when calling `transcribe_file` if needed.
47
+
48
+ ## Install SpeechBrain
49
+
50
+ First of all, please install SpeechBrain with the following command:
51
+
52
+ ```
53
+ pip install speechbrain
54
+ ```
55
+
56
+ Please notice that we encourage you to read our tutorials and learn more about
57
+ [SpeechBrain](https://speechbrain.github.io).
58
+
59
+ ## Transcribing Your Own Audio Files
60
+
61
+ ```python
62
+ from speechbrain.pretrained import EncoderDecoderASR
63
+ asr_model = EncoderDecoderASR.from_hparams(source="speechbrain/asr-transformer-switchboard", savedir="pretrained_models/asr-transformer-switchboard")
64
+ asr_model.transcribe_file("speechbrain/asr-transformer-switchboard/example.wav")
65
+ ```
66
+
67
+ ## Inference on GPU
68
+
69
+ To perform inference on the GPU, add `run_opts={"device":"cuda"}` when calling the `from_hparams` method.
70
+
71
+ ## Parallel Inference on a Batch
72
+
73
+ Please, [see this Colab notebook](https://colab.research.google.com/drive/1hX5ZI9S4jHIjahFCZnhwwQmFoGAi3tmu?usp=sharing) to figure out how to transcribe in parallel a batch of input sentences using a pre-trained model.
74
+
75
+ ## Training
76
+
77
+ The model was trained with SpeechBrain (commit hash: `70904d0`).
78
+ To train it from scratch follow these steps:
79
+
80
+ 1. Clone SpeechBrain:
81
+ ```bash
82
+ git clone https://github.com/speechbrain/speechbrain/
83
+ ```
84
+
85
+ 2. Install it:
86
+ ```bash
87
+ cd speechbrain
88
+ pip install -r requirements.txt
89
+ pip install -e .
90
+ ```
91
+
92
+ 3. Run Training:
93
+ ```bash
94
+ cd recipes/Switchboard/ASR/transformer
95
+ python train.py hparams/transformer.yaml --data_folder=your_data_folder
96
+ ```
97
+
98
+ ## Limitations
99
+
100
+ The SpeechBrain team does not provide any warranty on the performance achieved by this model when used on other datasets.
101
+
102
+ ## Credits
103
+
104
+ This model was trained with resources provided by the [THN Center for AI](https://www.th-nuernberg.de/en/kiz).
105
+
106
+ # About SpeechBrain
107
+
108
+ SpeechBrain is an open-source and all-in-one speech toolkit. It is designed to be simple, extremely flexible, and user-friendly.
109
+ Competitive or state-of-the-art performance is obtained in various domains.
110
+
111
+ - Website: https://speechbrain.github.io/
112
+ - GitHub: https://github.com/speechbrain/speechbrain/
113
+ - HuggingFace: https://huggingface.co/speechbrain/
114
+
115
+ # Citing SpeechBrain
116
+
117
+ Please cite SpeechBrain if you use it for your research or business.
118
+
119
+ ```bibtex
120
+ @misc{speechbrain,
121
+ title={{SpeechBrain}: A General-Purpose Speech Toolkit},
122
+ author={Mirco Ravanelli and Titouan Parcollet and Peter Plantinga and Aku Rouhe and Samuele Cornell and Loren Lugosch and Cem Subakan and Nauman Dawalatabad and Abdelwahab Heba and Jianyuan Zhong and Ju-Chieh Chou and Sung-Lin Yeh and Szu-Wei Fu and Chien-Feng Liao and Elena Rastorgueva and François Grondin and William Aris and Hwidong Na and Yan Gao and Renato De Mori and Yoshua Bengio},
123
+ year={2021},
124
+ eprint={2106.04624},
125
+ archivePrefix={arXiv},
126
+ primaryClass={eess.AS},
127
+ note={arXiv:2106.04624}
128
+ }
129
+ ```
asr.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ac0c043fd86ac73bb59ae43450b1a9df4a172eb4ca5af398d3e6e32e95410f02
3
+ size 111661649
config.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ {
2
+ "speechbrain_interface": "EncoderDecoderASR"
3
+ }
example.wav ADDED
Binary file (925 kB). View file
 
hyperparams.yaml ADDED
@@ -0,0 +1,156 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # ############################################################################
2
+ # Model: E2E ASR with Transformer
3
+ # Encoder: Transformer Encoder
4
+ # Decoder: Transformer Decoder + (CTC/ATT joint) beamsearch + TransformerLM
5
+ # Tokens: unigram
6
+ # losses: CTC + KLdiv (Label Smoothing loss)
7
+ # Training: Switchboard
8
+ # Authors: Jianyuan Zhong, Titouan Parcollet, Samuele Cornell, Dominik Wagner
9
+ # ############################################################################
10
+
11
+ # Feature parameters
12
+ sample_rate: 16000
13
+ n_fft: 400
14
+ n_mels: 80
15
+
16
+ ####################### Model parameters ###########################
17
+ # Transformer
18
+ transformer_input_size: 1280
19
+ d_model: 256
20
+ nhead: 4
21
+ num_encoder_layers: 12
22
+ num_decoder_layers: 6
23
+ d_ffn: 2048
24
+ transformer_dropout: 0.1
25
+ activation: !name:torch.nn.GELU
26
+ output_neurons: 2000
27
+
28
+ # Outputs
29
+ blank_index: 0
30
+ label_smoothing: 0.1
31
+ pad_index: 0
32
+ bos_index: 1
33
+ eos_index: 2
34
+ # unk_index: 0
35
+
36
+ # Decoding parameters
37
+ min_decode_ratio: 0.0
38
+ max_decode_ratio: 1.0
39
+ valid_search_interval: 10
40
+ valid_beam_size: 10
41
+ lm_weight: 0.30
42
+ test_beam_size: 60
43
+ ctc_weight_decode: 0.30
44
+ temperature: 1.0
45
+ temperature_lm: 1.0
46
+ using_eos_threshold: False
47
+ eos_threshold: 1.5
48
+ length_normalization: True
49
+ using_max_attn_shift: False
50
+ max_attn_shift: 30
51
+
52
+ CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
53
+ input_shape: (8, 10, 80)
54
+ num_blocks: 3
55
+ num_layers_per_block: 1
56
+ out_channels: (64, 64, 64)
57
+ kernel_sizes: (5, 5, 1)
58
+ strides: (2, 2, 1)
59
+ residuals: (False, False, True)
60
+
61
+ Transformer: !new:speechbrain.lobes.models.transformer.TransformerASR.TransformerASR # yamllint disable-line rule:line-length
62
+ input_size: !ref <transformer_input_size>
63
+ tgt_vocab: !ref <output_neurons>
64
+ d_model: !ref <d_model>
65
+ nhead: !ref <nhead>
66
+ num_encoder_layers: !ref <num_encoder_layers>
67
+ num_decoder_layers: !ref <num_decoder_layers>
68
+ d_ffn: !ref <d_ffn>
69
+ dropout: !ref <transformer_dropout>
70
+ activation: !ref <activation>
71
+ encoder_module: transformer
72
+ attention_type: regularMHA
73
+ normalize_before: True
74
+ causal: False
75
+
76
+ lm_model: !new:speechbrain.lobes.models.transformer.TransformerLM.TransformerLM # yamllint disable-line rule:line-length
77
+ vocab: !ref <output_neurons>
78
+ d_model: 264
79
+ d_embedding: 128
80
+ nhead: 12
81
+ num_encoder_layers: 12
82
+ num_decoder_layers: 0
83
+ d_ffn: 1024
84
+ dropout: 0.1
85
+ activation: !name:torch.nn.ReLU
86
+ normalize_before: False
87
+
88
+ tokenizer: !new:sentencepiece.SentencePieceProcessor
89
+
90
+ ctc_lin: !new:speechbrain.nnet.linear.Linear
91
+ input_size: !ref <d_model>
92
+ n_neurons: !ref <output_neurons>
93
+
94
+ seq_lin: !new:speechbrain.nnet.linear.Linear
95
+ input_size: !ref <d_model>
96
+ n_neurons: !ref <output_neurons>
97
+
98
+ asr_model: !new:torch.nn.ModuleList
99
+ - [!ref <CNN>, !ref <Transformer>, !ref <seq_lin>, !ref <ctc_lin>]
100
+
101
+ log_softmax: !new:torch.nn.LogSoftmax
102
+ dim: -1
103
+
104
+ normalizer: !new:speechbrain.processing.features.InputNormalization
105
+ norm_type: global
106
+
107
+ compute_features: !new:speechbrain.lobes.features.Fbank
108
+ sample_rate: !ref <sample_rate>
109
+ n_fft: !ref <n_fft>
110
+ n_mels: !ref <n_mels>
111
+
112
+ Tencoder: !new:speechbrain.lobes.models.transformer.TransformerASR.EncoderWrapper
113
+ transformer: !ref <Transformer>
114
+
115
+ encoder: !new:speechbrain.nnet.containers.LengthsCapableSequential
116
+ input_shape: [null, null, !ref <n_mels>]
117
+ compute_features: !ref <compute_features>
118
+ normalize: !ref <normalizer>
119
+ cnn: !ref <CNN>
120
+ transformer_encoder: !ref <Tencoder>
121
+
122
+ decoder: !new:speechbrain.decoders.S2STransformerBeamSearch
123
+ modules: [!ref <Transformer>, !ref <seq_lin>, !ref <ctc_lin>]
124
+ bos_index: !ref <bos_index>
125
+ eos_index: !ref <eos_index>
126
+ blank_index: !ref <blank_index>
127
+ min_decode_ratio: !ref <min_decode_ratio>
128
+ max_decode_ratio: !ref <max_decode_ratio>
129
+ beam_size: !ref <test_beam_size>
130
+ ctc_weight: !ref <ctc_weight_decode>
131
+ lm_weight: !ref <lm_weight>
132
+ lm_modules: !ref <lm_model>
133
+ temperature: !ref <temperature>
134
+ temperature_lm: !ref <temperature_lm>
135
+ using_eos_threshold: !ref <using_eos_threshold>
136
+ eos_threshold: !ref <eos_threshold>
137
+ length_normalization: !ref <length_normalization>
138
+ using_max_attn_shift: !ref <using_max_attn_shift>
139
+ max_attn_shift: !ref <max_attn_shift>
140
+
141
+ modules:
142
+ compute_features: !ref <compute_features>
143
+ normalizer: !ref <normalizer>
144
+ pre_transformer: !ref <CNN>
145
+ transformer: !ref <Transformer>
146
+ asr_model: !ref <asr_model>
147
+ lm_model: !ref <lm_model>
148
+ encoder: !ref <encoder>
149
+ decoder: !ref <decoder>
150
+
151
+ pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
152
+ loadables:
153
+ normalizer: !ref <normalizer>
154
+ asr: !ref <asr_model>
155
+ lm: !ref <lm_model>
156
+ tokenizer: !ref <tokenizer>
lm.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4e52dfb1d4d23ae2e861feffb04eadd94c8e34537c74633263966294becaa6ab
3
+ size 45759305
normalizer.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4537250f5851871c7eb0871bf72ef32a0096bae22d43f120ff6cf0a63c2443db
3
+ size 1703
tokenizer.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:25f80ce1b720439d9ebc46a6d3aa399bc2018eaf001084f6f63dc4b7e35f25d0
3
+ size 270824