Automatic Speech Recognition
NeMo
PyTorch
English
speech
audio
FastConformer
Conformer
NeMo
hf-asr-leaderboard
ctc
Eval Results
nithinraok commited on
Commit
e9993e4
1 Parent(s): 5e04eb8

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +293 -0
README.md CHANGED
@@ -1,3 +1,296 @@
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
15
+ - europarl
16
+ - multilingual_librispeech
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
+ - Conformer
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: Parakeet_XXL
38
+ results:
39
+ - task:
40
+ name: Automatic Speech Recognition
41
+ type: automatic-speech-recognition
42
+ dataset:
43
+ name: AMI (Meetings test)
44
+ type: edinburghcstr/ami
45
+ config: ihm
46
+ split: test
47
+ args:
48
+ language: en
49
+ metrics:
50
+ - name: Test WER
51
+ type: wer
52
+ value: 15.62
53
+ - task:
54
+ name: Automatic Speech Recognition
55
+ type: automatic-speech-recognition
56
+ dataset:
57
+ name: Earnings-22
58
+ type: revdotcom/earnings22
59
+ split: test
60
+ args:
61
+ language: en
62
+ metrics:
63
+ - name: Test WER
64
+ type: wer
65
+ value: 13.69
66
+ - task:
67
+ name: Automatic Speech Recognition
68
+ type: automatic-speech-recognition
69
+ dataset:
70
+ name: GigaSpeech
71
+ type: speechcolab/gigaspeech
72
+ split: test
73
+ args:
74
+ language: en
75
+ metrics:
76
+ - name: Test WER
77
+ type: wer
78
+ value: 10.27
79
+ - task:
80
+ name: Automatic Speech Recognition
81
+ type: automatic-speech-recognition
82
+ dataset:
83
+ name: LibriSpeech (clean)
84
+ type: librispeech_asr
85
+ config: other
86
+ split: test
87
+ args:
88
+ language: en
89
+ metrics:
90
+ - name: Test WER
91
+ type: wer
92
+ value: 1.83
93
+ - task:
94
+ name: Automatic Speech Recognition
95
+ type: automatic-speech-recognition
96
+ dataset:
97
+ name: LibriSpeech (other)
98
+ type: librispeech_asr
99
+ config: other
100
+ split: test
101
+ args:
102
+ language: en
103
+ metrics:
104
+ - name: Test WER
105
+ type: wer
106
+ value: 3.54
107
+ - task:
108
+ type: Automatic Speech Recognition
109
+ name: automatic-speech-recognition
110
+ dataset:
111
+ name: SPGI Speech
112
+ type: kensho/spgispeech
113
+ config: test
114
+ split: test
115
+ args:
116
+ language: en
117
+ metrics:
118
+ - name: Test WER
119
+ type: wer
120
+ value: 4.20
121
+ - task:
122
+ type: Automatic Speech Recognition
123
+ name: automatic-speech-recognition
124
+ dataset:
125
+ name: tedlium-v3
126
+ type: LIUM/tedlium
127
+ config: release1
128
+ split: test
129
+ args:
130
+ language: en
131
+ metrics:
132
+ - name: Test WER
133
+ type: wer
134
+ value: 3.54
135
+ - task:
136
+ name: Automatic Speech Recognition
137
+ type: automatic-speech-recognition
138
+ dataset:
139
+ name: Vox Populi
140
+ type: facebook/voxpopuli
141
+ config: en
142
+ split: test
143
+ args:
144
+ language: en
145
+ metrics:
146
+ - name: Test WER
147
+ type: wer
148
+ value: 6.53
149
+ - task:
150
+ type: Automatic Speech Recognition
151
+ name: automatic-speech-recognition
152
+ dataset:
153
+ name: Mozilla Common Voice 9.0
154
+ type: mozilla-foundation/common_voice_9_0
155
+ config: en
156
+ split: test
157
+ args:
158
+ language: en
159
+ metrics:
160
+ - name: Test WER
161
+ type: wer
162
+ value: 9.02
163
+
164
+ metrics:
165
+ - wer
166
+ pipeline_tag: automatic-speech-recognition
167
  ---
168
+
169
+ # Parakeet CTC 1.1B (en)
170
+
171
+ <style>
172
+ img {
173
+ display: inline;
174
+ }
175
+ </style>
176
+
177
+ [![Model architecture](https://img.shields.io/badge/Model_Arch-FastConformer--CTC-lightgrey#model-badge)](#model-architecture)
178
+ | [![Model size](https://img.shields.io/badge/Params-1.1B-lightgrey#model-badge)](#model-architecture)
179
+ | [![Language](https://img.shields.io/badge/Language-en-lightgrey#model-badge)](#datasets)
180
+
181
+
182
+ parakeet-rnnt-1.1b is an ASR model that transcribes speech in lower case English alphabet. This model is jointly developed by [NVIDIA NeMo](https://github.com/NVIDIA/NeMo) and [Suno.ai](https://www.suno.ai/) teams.
183
+ It is an XXL version of FastConformer CTC [1] (around 1.1B parameters) model.
184
+ 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.
185
+
186
+ ## NVIDIA NeMo: Training
187
+
188
+ 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.
189
+ ```
190
+ pip install nemo_toolkit['all']
191
+ ```
192
+
193
+ ## How to Use this Model
194
+
195
+ 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.
196
+
197
+ ### Automatically instantiate the model
198
+
199
+ ```python
200
+ import nemo.collections.asr as nemo_asr
201
+ asr_model = nemo_asr.models.EncDecRNNTBPEModel.from_pretrained(model_name="nvidia/parakeet-ctc-1.1b")
202
+ ```
203
+
204
+ ### Transcribing using Python
205
+ First, let's get a sample
206
+ ```
207
+ wget https://dldata-public.s3.us-east-2.amazonaws.com/2086-149220-0033.wav
208
+ ```
209
+ Then simply do:
210
+ ```
211
+ asr_model.transcribe(['2086-149220-0033.wav'])
212
+ ```
213
+
214
+ ### Transcribing many audio files
215
+
216
+ ```shell
217
+ python [NEMO_GIT_FOLDER]/examples/asr/transcribe_speech.py
218
+ pretrained_name="nvidia/parakeet-ctc-1.1b"
219
+ audio_dir="<DIRECTORY CONTAINING AUDIO FILES>"
220
+ ```
221
+
222
+ ### Input
223
+
224
+ This model accepts 16000 Hz mono-channel audio (wav files) as input.
225
+
226
+ ### Output
227
+
228
+ This model provides transcribed speech as a string for a given audio sample.
229
+
230
+ ## Model Architecture
231
+
232
+ FastConformer [1] is an optimized version of the Conformer model with 8x depthwise-separable convolutional downsampling. The model is trained using CTC 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).
233
+
234
+ ## Training
235
+
236
+ 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_ctc/speech_to_text_ctc_bpe.py) and this [base config](https://github.com/NVIDIA/NeMo/blob/main/examples/asr/conf/fastconformer/fast-conformer_ctc_bpe.yaml).
237
+
238
+ 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).
239
+
240
+ ### Datasets
241
+
242
+ The model was trained on 65K hours of English speech collected and prepared by NVIDIA NeMo and Suno teams.
243
+
244
+ The training dataset consists of private subset with 40K hours of English speech plus 25K hours from the following public datasets:
245
+
246
+ - Librispeech 960 hours of English speech
247
+ - Fisher Corpus
248
+ - Switchboard-1 Dataset
249
+ - WSJ-0 and WSJ-1
250
+ - National Speech Corpus (Part 1, Part 6)
251
+ - VCTK
252
+ - VoxPopuli (EN)
253
+ - Europarl-ASR (EN)
254
+ - Multilingual Librispeech (MLS EN) - 2,000 hour subset
255
+ - Mozilla Common Voice (v7.0)
256
+ - People's Speech - 12,000 hour subset
257
+
258
+ ## Performance
259
+
260
+ 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.
261
+
262
+ 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.
263
+
264
+ |**Version**|**Tokenizer**|**Vocabulary Size**|**AMI**|**Earnings-22**|**Giga Speech**|**LS test-clean**|**SPGI Speech**|**TEDLIUM-v3**|**Vox Populi**|**Common Voice**|
265
+ |---------|-----------------------|-----------------|---------------|---------------|------------|-----------|-----|-------|------|------|
266
+ | 1.22.0 | SentencePiece Unigram | 1024 | 15.62 | 13.69 | 10.27 | 1.83 | 3.54 | 4.20 | 3.54 | 6.53 | 9.02 |
267
+
268
+ These are greedy WER numbers without external LM. More details on evaluation can be found at [HuggingFace ASR Leaderboard](https://huggingface.co/spaces/hf-audio/open_asr_leaderboard)
269
+
270
+ ## NVIDIA Riva: Deployment
271
+
272
+ [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.
273
+ Additionally, Riva provides:
274
+
275
+ * 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
276
+ * 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
277
+ * Streaming speech recognition, Kubernetes compatible scaling, and enterprise-grade support
278
+
279
+ Although this model isn’t supported yet by Riva, the [list of supported models is here](https://huggingface.co/models?other=Riva).
280
+ Check out [Riva live demo](https://developer.nvidia.com/riva#demos).
281
+
282
+ ## References
283
+ [1] [Fast Conformer with Linearly Scalable Attention for Efficient Speech Recognition](https://arxiv.org/abs/2305.05084)
284
+
285
+ [2] [Google Sentencepiece Tokenizer](https://github.com/google/sentencepiece)
286
+
287
+ [3] [NVIDIA NeMo Toolkit](https://github.com/NVIDIA/NeMo)
288
+
289
+ [4] [Suno.ai](https://suno.ai/)
290
+
291
+ [5] [HuggingFace ASR Leaderboard](https://huggingface.co/spaces/hf-audio/open_asr_leaderboard)
292
+
293
+
294
+ ## Licence
295
+
296
+ 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.