franjamonga commited on
Commit
6fceb40
1 Parent(s): 0feb365

Upload config

Browse files
Files changed (1) hide show
  1. README.md +143 -0
README.md ADDED
@@ -0,0 +1,143 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - en
4
+ license: cc-by-4.0
5
+ library_name: nemo
6
+ datasets:
7
+ - librispeech_asr
8
+ - VOXCCELEB-1
9
+ - VOXCCELEB-2
10
+ - FISHER
11
+ - Switchboard
12
+ - SRE(2004-2010)
13
+ thumbnail: null
14
+ tags:
15
+ - speaker-recognition
16
+ - speech
17
+ - audio
18
+ - speaker-verification
19
+ - titanet
20
+ - speaker-diarization
21
+ - NeMo
22
+ - pytorch
23
+ model-index:
24
+ - name: speakerverification_en
25
+ results: []
26
+
27
+ ---
28
+
29
+
30
+ ## Model Overview
31
+
32
+ This model extracts speaker embeddings from given speech, which is the backbone for speaker verification and diarization tasks.
33
+ It is a "large" version of TitaNet (around 23M parameters) models.
34
+ See the [model architecture](#model-architecture) section and [NeMo documentation](https://docs.nvidia.com/deeplearning/nemo/user
35
+
36
+ ## How to Use this Model
37
+
38
+ 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.
39
+
40
+ ### Automatically instantiate the model
41
+
42
+ ```python
43
+ import nemo.collections.asr as nemo_asr
44
+ speaker_model = nemo_asr.models.EncDecSpeakerLabelModel.from_pretrained("nvidia/speakerverification_en_titanet_large")
45
+ ```
46
+
47
+ ### Embedding Extraction
48
+
49
+ Using
50
+
51
+ ```python
52
+ emb = speaker_model.get_embedding("an255-fash-b.wav")
53
+ ```
54
+
55
+ ### Verifying two utterances (Speaker Verification)
56
+
57
+ Now to check if two audio files are from the same speaker or not, simply do:
58
+
59
+ ```python
60
+ speaker_model.verify_speakers("an255-fash-b.wav","cen7-fash-b.wav")
61
+ ```
62
+
63
+ ### Extracting Embeddings for more audio files
64
+
65
+ To extract embeddings from a bunch of audio files:
66
+
67
+ Write audio files to a `manifest.json` file with lines as in format:
68
+
69
+ ```json
70
+ {"audio_filepath": "<absolute path to dataset>/audio_file.wav", "duration": "duration of file in sec", "label": "speaker_id"}
71
+ ```
72
+
73
+ Then running following script will extract embeddings and writes to current working directory:
74
+ ```shell
75
+ python <NeMo_root>/examples/speaker_tasks/recognition/extract_speaker_embeddings.py --manifest=manifest.json
76
+ ```
77
+
78
+ ### Input
79
+
80
+ This model accepts 16000 KHz Mono-channel Audio (wav files) as input.
81
+
82
+ ### Output
83
+
84
+ This model provides speaker embeddings for an audio file.
85
+
86
+ ## Model Architecture
87
+
88
+ TitaNet model is a depth-wise separable conv1D model [1] for Speaker Verification and diarization tasks. You may find more info on the detail of this model here: [TitaNet-Model](https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_recognition/models.html).
89
+
90
+ ## Training
91
+
92
+ 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/speaker_tasks/recognition/speaker_reco.py) and this [base config](https://github.com/NVIDIA/NeMo/blob/main/examples/speaker_tasks/recognition/conf/titanet-large.yaml).
93
+
94
+ ### Datasets
95
+
96
+ All the models in this collection are trained on a composite dataset comprising several thousand hours of English speech:
97
+
98
+ - Voxceleb-1
99
+ - Voxceleb-2
100
+ - Fisher
101
+ - Switchboard
102
+ - Librispeech
103
+ - SRE (2004-2010)
104
+
105
+
106
+ ## Performance
107
+
108
+ Performances of the these models are reported in terms of Equal Error Rate (EER%) on speaker verification evaluation trial files and as Diarization Error Rate (DER%) on diarization test sessions.
109
+
110
+ * Speaker Verification (EER%)
111
+ | Version | Model | Model Size | VoxCeleb1 (Cleaned trial file) |
112
+ |---------|--------------|-----|---------------|
113
+ | 1.10.0 | TitaNet-Large | 23M | 0.66 |
114
+
115
+ * Speaker Diarization (DER%)
116
+ | Version | Model | Model Size | Evaluation Condition | NIST SRE 2000 | AMI (Lapel) | AMI (MixHeadset) | CH109 |
117
+ |---------|--------------|-----|----------------------|---------------|-------------|------------------|-------|
118
+ | 1.10.0 | TitaNet-Large | 23M | Oracle VAD KNOWN # of Speakers | 6.73 | 2.03 | 1.73 | 1.19 |
119
+ | 1.10.0 | TitaNet-Large | 23M | Oracle VAD UNKNOWN # of Speakers | 5.38 | 2.03 | 1.89 | 1.63 |
120
+
121
+ ## Limitations
122
+ This model is trained on both telephonic and non-telephonic speech from voxceleb datasets, Fisher and switch board. If your domain of data differs from trained data or doesnot show relatively good performance consider finetuning for that speech domain.
123
+
124
+ ## NVIDIA Riva: Deployment
125
+
126
+ [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.
127
+ Additionally, Riva provides:
128
+
129
+ * 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
130
+ * 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
131
+ * Streaming speech recognition, Kubernetes compatible scaling, and enterprise-grade support
132
+
133
+ Although this model isn’t supported yet by Riva, the [list of supported models is here](https://huggingface.co/models?other=Riva).
134
+ Check out [Riva live demo](https://developer.nvidia.com/riva#demos).
135
+
136
+ ## References
137
+ [1] [TitaNet: Neural Model for Speaker Representation with 1D Depth-wise Separable convolutions and global context](https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9746806)
138
+ [2] [NVIDIA NeMo Toolkit](https://github.com/NVIDIA/NeMo)
139
+
140
+ ## Licence
141
+
142
+ 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.
143
+