nithinraok commited on
Commit
6e56ffd
1 Parent(s): 76f8530

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +255 -0
README.md CHANGED
@@ -1,3 +1,258 @@
1
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2
  license: cc-by-4.0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ language:
3
+ - en
4
+ library_name: nemo
5
+ datasets:
6
+ - librispeech_asr
7
+ - fisher_corpus
8
+ - Switchboard-1
9
+ - WSJ-0
10
+ - WSJ-1
11
+ - National-Singapore-Corpus-Part-1
12
+ - National-Singapore-Corpus-Part-6
13
+ - vctk
14
+ - VoxPopuli-(EN)
15
+ - Europarl-ASR-(EN)
16
+ - Multilingual-LibriSpeech-(2000-hours)
17
+ - mozilla-foundation/common_voice_8_0
18
+ - MLCommons/peoples_speech
19
+ thumbnail: null
20
+ tags:
21
+ - automatic-speech-recognition
22
+ - speech
23
+ - audio
24
+ - Transducer
25
+ - FastConformer
26
+ - Transformer
27
+ - pytorch
28
+ - NeMo
29
+ - hf-asr-leaderboard
30
  license: cc-by-4.0
31
+ widget:
32
+ - example_title: Librispeech sample 1
33
+ src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
34
+ - example_title: Librispeech sample 2
35
+ src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
36
+ model-index:
37
+ - name: stt_en_fastconformer_transducer_xlarge
38
+ results:
39
+ - task:
40
+ name: Automatic Speech Recognition
41
+ type: automatic-speech-recognition
42
+ dataset:
43
+ name: LibriSpeech (clean)
44
+ type: librispeech_asr
45
+ config: clean
46
+ split: test
47
+ args:
48
+ language: en
49
+ metrics:
50
+ - name: Test WER
51
+ type: wer
52
+ value: 1.64
53
+ - task:
54
+ name: Automatic Speech Recognition
55
+ type: automatic-speech-recognition
56
+ dataset:
57
+ name: LibriSpeech (other)
58
+ type: librispeech_asr
59
+ config: other
60
+ split: test
61
+ args:
62
+ language: en
63
+ metrics:
64
+ - name: Test WER
65
+ type: wer
66
+ value: 3.07
67
+ - task:
68
+ type: Automatic Speech Recognition
69
+ name: automatic-speech-recognition
70
+ dataset:
71
+ name: Multilingual LibriSpeech
72
+ type: facebook/multilingual_librispeech
73
+ config: english
74
+ split: test
75
+ args:
76
+ language: en
77
+ metrics:
78
+ - name: Test WER
79
+ type: wer
80
+ value: 4.90
81
+ - task:
82
+ type: Automatic Speech Recognition
83
+ name: automatic-speech-recognition
84
+ dataset:
85
+ name: Mozilla Common Voice 7.0
86
+ type: mozilla-foundation/common_voice_7_0
87
+ config: en
88
+ split: test
89
+ args:
90
+ language: en
91
+ metrics:
92
+ - name: Test WER
93
+ type: wer
94
+ value: 6.5
95
+ - task:
96
+ type: Automatic Speech Recognition
97
+ name: automatic-speech-recognition
98
+ dataset:
99
+ name: Wall Street Journal 92
100
+ type: wsj_0
101
+ args:
102
+ language: en
103
+ metrics:
104
+ - name: Test WER
105
+ type: wer
106
+ value: 1.11
107
+ - task:
108
+ type: Automatic Speech Recognition
109
+ name: automatic-speech-recognition
110
+ dataset:
111
+ name: Wall Street Journal 93
112
+ type: wsj_1
113
+ args:
114
+ language: en
115
+ metrics:
116
+ - name: Test WER
117
+ type: wer
118
+ value: 1.99
119
+ - task:
120
+ name: Automatic Speech Recognition
121
+ type: automatic-speech-recognition
122
+ dataset:
123
+ name: National Singapore Corpus
124
+ type: nsc_part_1
125
+ split: test
126
+ args:
127
+ language: en
128
+ metrics:
129
+ - name: Test WER
130
+ type: wer
131
+ value: 4.47
132
+
133
+
134
  ---
135
+
136
+ # NVIDIA FastConformer-Transducer XLarge (en)
137
+
138
+ <style>
139
+ img {
140
+ display: inline;
141
+ }
142
+ </style>
143
+
144
+ | [![Model architecture](https://img.shields.io/badge/Model_Arch-FastConformer--Transducer-lightgrey#model-badge)](#model-architecture)
145
+ | [![Model size](https://img.shields.io/badge/Params-618M-lightgrey#model-badge)](#model-architecture)
146
+ | [![Language](https://img.shields.io/badge/Language-en-lightgrey#model-badge)](#datasets)
147
+
148
+
149
+ This model transcribes speech in lower case English alphabet.
150
+ It is a "large" version of FastConformer Transducer (around 618M parameters) model.
151
+ See the [model architecture](#model-architecture) section and [NeMo documentation](https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/models.html#fast-conformer) for complete architecture details.
152
+
153
+ ## NVIDIA NeMo: Training
154
+
155
+ To train, fine-tune or play with the model you will need to install [NVIDIA NeMo](https://github.com/NVIDIA/NeMo). We recommend you install it after you've installed latest Pytorch version.
156
+ ```
157
+ pip install nemo_toolkit['all']
158
+ ```
159
+
160
+ ## How to Use this Model
161
+
162
+ The model is available for use in the NeMo toolkit [3], and can be used as a pre-trained checkpoint for inference or for fine-tuning on another dataset.
163
+
164
+ ### Automatically instantiate the model
165
+
166
+ ```python
167
+ import nemo.collections.asr as nemo_asr
168
+ asr_model = nemo_asr.models.EncDecRNNTBPEModel.from_pretrained(model_name="nvidia/stt_en_fastconformer_transducer_xlarge")
169
+ ```
170
+
171
+ ### Transcribing using Python
172
+ First, let's get a sample
173
+ ```
174
+ wget https://dldata-public.s3.us-east-2.amazonaws.com/2086-149220-0033.wav
175
+ ```
176
+ Then simply do:
177
+ ```
178
+ asr_model.transcribe(['2086-149220-0033.wav'])
179
+ ```
180
+
181
+ ### Transcribing many audio files
182
+
183
+ ```shell
184
+ python [NEMO_GIT_FOLDER]/examples/asr/transcribe_speech.py
185
+ pretrained_name="nvidia/stt_en_fastconformer_transducer_xlarge"
186
+ audio_dir="<DIRECTORY CONTAINING AUDIO FILES>"
187
+ ```
188
+
189
+ ### Input
190
+
191
+ This model accepts 16000 Hz Mono-channel Audio (wav files) as input.
192
+
193
+ ### Output
194
+
195
+ This model provides transcribed speech as a string for a given audio sample.
196
+
197
+ ## Model Architecture
198
+
199
+ FastConformer [1] is an optimized version of the Conformer model with 8x depthwise-separable convolutional downsampling. The model is trained in a multitask setup with a Transducer decoder (RNNT) loss. You may find more information on the details of FastConformer here: [Fast-Conformer Model](https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/models.html#fast-conformer).
200
+
201
+ ## Training
202
+
203
+ The NeMo toolkit [3] was used for training the models for over several hundred epochs. These model are trained with this [example script](https://github.com/NVIDIA/NeMo/blob/main/examples/asr/asr_transducer/speech_to_text_rnnt_bpe.py) and this [base config](https://github.com/NVIDIA/NeMo/blob/main/examples/asr/conf/fastconformer/fast-conformer_transducer_bpe.yaml).
204
+
205
+ The tokenizers for these models were built using the text transcripts of the train set with this [script](https://github.com/NVIDIA/NeMo/blob/main/scripts/tokenizers/process_asr_text_tokenizer.py).
206
+
207
+ ### Datasets
208
+
209
+ The model in this collection is trained on a composite dataset (NeMo ASRSet En) comprising several thousand hours of English speech:
210
+
211
+ - Librispeech 960 hours of English speech
212
+ - Fisher Corpus
213
+ - Switchboard-1 Dataset
214
+ - WSJ-0 and WSJ-1
215
+ - National Speech Corpus (Part 1, Part 6)
216
+ - VCTK
217
+ - VoxPopuli (EN)
218
+ - Europarl-ASR (EN)
219
+ - Multilingual Librispeech (MLS EN) - 2,000 hrs subset
220
+ - Mozilla Common Voice (v7.0)
221
+ - People's Speech - 12,000 hrs subset
222
+
223
+ ## Performance
224
+
225
+ The performance of Automatic Speech Recognition models is measuring using Word Error Rate. Since this dataset is trained on multiple domains and a much larger corpus, it will generally perform better at transcribing audio in general.
226
+
227
+ The following tables summarizes the performance of the available models in this collection with the Transducer decoder. Performances of the ASR models are reported in terms of Word Error Rate (WER%) with greedy decoding.
228
+
229
+ |**Version**|**Tokenizer**|**Vocabulary Size**|**LS test-other**|**LS test-clean**|**WSJ Eval92**|**WSJ Dev93**|**NSC Part 1**|**MLS Test**|**MCV Test 7.0**| Train Dataset |
230
+ |---------|-----------------------|-----------------|---------------|---------------|------------|-----------|-----|-------|------|------|
231
+ | 1.18.0 | SentencePiece Unigram | 1024 | 3.07 | 1.64 | 1.11 | 1.99 | 4.47 | 4.90 | 6.5 | NeMo ASRSET 3.0 |
232
+
233
+
234
+ ## Limitations
235
+ Since this model was trained on publically available speech datasets, the performance of this model might degrade for speech which includes technical terms, or vernacular that the model has not been trained on. The model might also perform worse for accented speech.
236
+
237
+ ## NVIDIA Riva: Deployment
238
+
239
+ [NVIDIA Riva](https://developer.nvidia.com/riva), is an accelerated speech AI SDK deployable on-prem, in all clouds, multi-cloud, hybrid, on edge, and embedded.
240
+ Additionally, Riva provides:
241
+
242
+ * World-class out-of-the-box accuracy for the most common languages with model checkpoints trained on proprietary data with hundreds of thousands of GPU-compute hours
243
+ * Best in class accuracy with run-time word boosting (e.g., brand and product names) and customization of acoustic model, language model, and inverse text normalization
244
+ * Streaming speech recognition, Kubernetes compatible scaling, and enterprise-grade support
245
+
246
+ Although this model isn’t supported yet by Riva, the [list of supported models is here](https://huggingface.co/models?other=Riva).
247
+ Check out [Riva live demo](https://developer.nvidia.com/riva#demos).
248
+
249
+ ## References
250
+ [1] [Fast Conformer with Linearly Scalable Attention for Efficient Speech Recognition](https://arxiv.org/abs/2305.05084)
251
+
252
+ [2] [Google Sentencepiece Tokenizer](https://github.com/google/sentencepiece)
253
+
254
+ [3] [NVIDIA NeMo Toolkit](https://github.com/NVIDIA/NeMo)
255
+
256
+ ## Licence
257
+
258
+ License to use this model is covered by the [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/). By downloading the public and release version of the model, you accept the terms and conditions of the [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/) license.