vnoroozi commited on
Commit
284dd9e
1 Parent(s): a02cb3a

added the model card.

Browse files
Files changed (1) hide show
  1. README.md +167 -0
README.md CHANGED
@@ -1,3 +1,170 @@
1
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2
  license: cc-by-4.0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ language:
3
+ - en
4
+ library_name: nemo
5
+ datasets:
6
+ - AISHELL-2
7
+ thumbnail: null
8
+ tags:
9
+ - automatic-speech-recognition
10
+ - speech
11
+ - audio
12
+ - Transducer
13
+ - Conformer
14
+ - Transformer
15
+ - pytorch
16
+ - NeMo
17
+ - hf-asr-leaderboard
18
  license: cc-by-4.0
19
+ model-index:
20
+ - name: stt_zh_conformer_transducer_large
21
+ results:
22
+ - task:
23
+ name: Automatic Speech Recognition
24
+ type: automatic-speech-recognition
25
+ dataset:
26
+ name: AISHELL-2 Test IOS
27
+ type: aishell2_test_ios
28
+ config: Mandarin
29
+ split: test
30
+ args:
31
+ language: zh
32
+ metrics:
33
+ - name: Test WER
34
+ type: wer
35
+ value: 5.3
36
+ - task:
37
+ type: Automatic Speech Recognition
38
+ name: automatic-speech-recognition
39
+ dataset:
40
+ name: AISHELL-2 Test Android
41
+ type: aishell2_test_android
42
+ config: Mandarin
43
+ split: test
44
+ args:
45
+ language: zh
46
+ metrics:
47
+ - name: Test WER
48
+ type: wer
49
+ value: 5.7
50
+ - task:
51
+ type: Automatic Speech Recognition
52
+ name: automatic-speech-recognition
53
+ dataset:
54
+ name: AISHELL-2 Test Mic
55
+ type: aishell2_test_mic
56
+ config: Mandarin
57
+ split: test
58
+ args:
59
+ language: zh
60
+ metrics:
61
+ - name: Test WER
62
+ type: wer
63
+ value: 5.6
64
  ---
65
+
66
+ # NVIDIA Conformer-Transducer Large (zh-ZH)
67
+
68
+ <style>
69
+ img {
70
+ display: inline;
71
+ }
72
+ </style>
73
+
74
+ | [![Model architecture](https://img.shields.io/badge/Model_Arch-Conformer--Transducer-lightgrey#model-badge)](#model-architecture)
75
+ | [![Model size](https://img.shields.io/badge/Params-120M-lightgrey#model-badge)](#model-architecture)
76
+ | [![Language](https://img.shields.io/badge/Language-zh--ZH-lightgrey#model-badge)](#datasets)
77
+
78
+
79
+ This model transcribes speech in Mandarin alphabet.
80
+ It is a large version of Conformer-Transducer (around 120M parameters) model.
81
+ See the [model architecture](#model-architecture) section and [NeMo documentation](https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/models.html#conformer-transducer) for complete architecture details.
82
+
83
+ ## NVIDIA NeMo: Training
84
+
85
+ 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.
86
+ ```
87
+ pip install nemo_toolkit['all']
88
+ ```
89
+
90
+ ## How to Use this Model
91
+
92
+ 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.
93
+
94
+ ### Automatically instantiate the model
95
+
96
+ ```python
97
+ import nemo.collections.asr as nemo_asr
98
+ asr_model = nemo_asr.models.EncDecRNNTBPEModel.from_pretrained("nvidia/stt_zh_conformer_transducer_large")
99
+ ```
100
+
101
+ ### Transcribing using Python
102
+ First, let's get a sample
103
+ ```
104
+ wget https://dldata-public.s3.us-east-2.amazonaws.com/2086-149220-0033.wav
105
+ ```
106
+ Then simply do:
107
+ ```
108
+ asr_model.transcribe(['2086-149220-0033.wav'])
109
+ ```
110
+
111
+ ### Transcribing many audio files
112
+
113
+ ```shell
114
+ python [NEMO_GIT_FOLDER]/examples/asr/transcribe_speech.py
115
+ pretrained_name="nvidia/stt_zh_conformer_transducer_large"
116
+ audio_dir="<DIRECTORY CONTAINING AUDIO FILES>"
117
+ ```
118
+
119
+ ### Input
120
+
121
+ This model accepts 16000 KHz Mono-channel Audio (wav files) as input.
122
+
123
+ ### Output
124
+
125
+ This model provides transcribed speech as a string for a given audio sample.
126
+
127
+ ## Model Architecture
128
+
129
+ Conformer-Transducer model is an autoregressive variant of Conformer model [1] for Automatic Speech Recognition which uses Transducer loss/decoding instead of CTC Loss. You may find more info on the detail of this model here: [Conformer-Transducer Model](https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/models.html).
130
+
131
+ ## Training
132
+
133
+ 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/conformer/conformer_transducer_bpe.yaml).
134
+
135
+ ### Datasets
136
+
137
+ All the models in this collection are trained on AISHELL2 [4] comprising of Mandarin speech:
138
+
139
+ ## Performance
140
+
141
+ The list of the available models in this collection is shown in the following table. Performances of the ASR models are reported in terms of Word Error Rate (WER%) with greedy decoding.
142
+
143
+ | Version | Tokenizer | Vocabulary Size | AISHELL2 Test IOS | AISHELL2 Test Android | AISHELL2 Test Mic | Train Dataset |
144
+ |---------|-----------|-----------------|-------------------|-----------------------|-------------------|---------------|
145
+ | 1.10.0 | Characters| 1024 | 5.3 | 5.7 | 5.6 | AISHELL-2 |
146
+
147
+ ## Limitations
148
+ Since this model was trained on publicly 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.
149
+
150
+ ## NVIDIA Riva: Deployment
151
+
152
+ [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.
153
+ Additionally, Riva provides:
154
+
155
+ * 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
156
+ * 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
157
+ * Streaming speech recognition, Kubernetes compatible scaling, and enterprise-grade support
158
+
159
+ Although this model isn’t supported yet by Riva, the [list of supported models is here](https://huggingface.co/models?other=Riva).
160
+ Check out [Riva live demo](https://developer.nvidia.com/riva#demos).
161
+
162
+ ## References
163
+ [1] [Conformer: Convolution-augmented Transformer for Speech Recognition](https://arxiv.org/abs/2005.08100)
164
+ [2] [Google Sentencepiece Tokenizer](https://github.com/google/sentencepiece)
165
+ [3] [NVIDIA NeMo Toolkit](https://github.com/NVIDIA/NeMo)
166
+ [4] [AISHELL-2: Transforming Mandarin ASR Research Into Industrial Scale](https://arxiv.org/abs/1808.10583)
167
+
168
+ ## Licence
169
+
170
+ 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.