nguyenvulebinh commited on
Commit
d1389bc
·
verified ·
1 Parent(s): 6353b9e

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +155 -0
README.md ADDED
@@ -0,0 +1,155 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ {}
3
+ ---
4
+
5
+ **This is the dataset for training MSA-ASR model**
6
+
7
+ # MSA-ASR
8
+
9
+ Multilingual Speaker-Attributed Automatic Speech Recognition
10
+
11
+ ### Demo
12
+
13
+ <video src="https://huggingface.co/nguyenvulebinh/MSA-ASR/resolve/main/demo_sa-asr.mp4" width="640" height="480" controls></video>
14
+
15
+ ### Introduction
16
+
17
+ This repository provides an implementation of a Speaker-Attributed Automatic Speech Recognition model. The model performs both multilingual speech recognition and speaker embedding extraction, enabling speaker differentiation.
18
+
19
+ Model architecture
20
+
21
+ ![MSA-ASR Model](https://github.com/nguyenvulebinh/MSA-ASR/blob/main/resource/model.png?raw=true)
22
+
23
+
24
+ ### Setup
25
+
26
+ ```
27
+ git clone [email protected]:nguyenvulebinh/MSA-ASR.git
28
+ cd MSA-ASR
29
+ conda create -n MSA-ASR python=3.10
30
+ conda activate MSA-ASR
31
+ pip install -r requirements.txt
32
+ ```
33
+
34
+ Test script:
35
+
36
+ ```
37
+ python infer.py
38
+ ```
39
+
40
+ ### Training Dataset
41
+
42
+ *From ASR to SA-ASR dataset:*
43
+
44
+ - Segment ASR data into single-speaker turns.
45
+ - Match turns into group which may come from the same speaker by using speaker embedding cosine similarity.
46
+ - Pick a few groups, each group a few turns.
47
+ - Concatenate turns in random order.
48
+
49
+ ![MSA-ASR Dataset](https://github.com/nguyenvulebinh/MSA-ASR/blob/main/resource/sa_asr_data_pipeline.png?raw=true)
50
+
51
+ *In total:*
52
+
53
+ - 15.5M turns
54
+ - 14k audio hours
55
+ - English only
56
+
57
+ Dataset is openly available in [HF Dataset](https://huggingface.co/datasets/nguyenvulebinh/spk-attribute)
58
+
59
+ *Example*
60
+
61
+ Audio
62
+
63
+ <audio controls>
64
+ <source src="https://huggingface.co/nguyenvulebinh/MSA-ASR/resolve/main/sample_augment.wav" type="audio/wav">
65
+ Your browser does not support the audio element.
66
+ </audio>
67
+
68
+
69
+ Label:
70
+
71
+ ```code
72
+ spk_1 A 0.00 1.58 »spk_1
73
+ spk_1 A 0.00 1.58 Pacifica
74
+ spk_1 A 1.58 0.68 continues
75
+ spk_1 A 2.27 0.52 today
76
+ spk_1 A 2.79 0.24 to
77
+ spk_1 A 3.03 0.20 be
78
+ spk_1 A 3.23 0.14 a
79
+ spk_1 A 3.37 0.54 listener
80
+ spk_1 A 3.91 0.80 supported
81
+ spk_1 A 4.71 0.70 network
82
+ spk_1 A 5.42 0.38 of
83
+ spk_2 A 5.80 0.12 »spk_2
84
+ spk_2 A 5.80 0.12 At
85
+ spk_2 A 5.92 0.42 home,
86
+ spk_2 A 6.34 0.18 an
87
+ spk_2 A 6.52 0.38 Aed
88
+ spk_2 A 6.90 0.26 is
89
+ spk_2 A 7.16 0.18 an
90
+ spk_2 A 7.34 0.56 automated
91
+ spk_2 A 7.90 0.60 external
92
+ spk_2 A 8.50 0.90 defibrillator.
93
+ spk_2 A 9.40 0.40 It's
94
+ spk_2 A 9.81 0.08 the
95
+ spk_2 A 9.89 0.36 device
96
+ spk_2 A 10.25 0.08 you
97
+ spk_2 A 10.33 0.16 use
98
+ spk_2 A 10.49 0.12 when
99
+ spk_2 A 10.61 0.10 your
100
+ spk_2 A 10.73 0.16 heart
101
+ spk_2 A 10.89 0.18 goes
102
+ spk_2 A 11.07 0.12 into
103
+ spk_2 A 11.19 0.38 cardiac
104
+ spk_2 A 11.57 0.38 arrest
105
+ spk_2 A 11.95 0.18 to
106
+ spk_2 A 12.13 0.36 shock
107
+ spk_2 A 12.49 0.14 it
108
+ spk_2 A 12.63 0.28 back
109
+ spk_2 A 12.91 0.22 into
110
+ spk_2 A 13.13 0.06 a
111
+ spk_2 A 13.19 0.32 normal
112
+ spk_2 A 13.51 0.88 rhythm.
113
+ spk_1 A 14.40 1.38 »spk_1
114
+ spk_1 A 14.40 1.38 stations.
115
+ ```
116
+
117
+ ### Citation
118
+
119
+ ```bibtex
120
+ @INPROCEEDINGS{10889116,
121
+ author={Nguyen, Thai-Binh and Waibel, Alexander},
122
+ booktitle={ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
123
+ title={MSA-ASR: Efficient Multilingual Speaker Attribution with frozen ASR Models},
124
+ year={2025},
125
+ volume={},
126
+ number={},
127
+ pages={1-5},
128
+ keywords={Training;Adaptation models;Limiting;Predictive models;Data models;Robustness;Multilingual;Data mining;Speech processing;Standards;speaker-attributed;asr;multilingual},
129
+ doi={10.1109/ICASSP49660.2025.10889116}}
130
+
131
+ @INPROCEEDINGS{10446589,
132
+ author={Nguyen, Thai-Binh and Waibel, Alexander},
133
+ booktitle={ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
134
+ title={Synthetic Conversations Improve Multi-Talker ASR},
135
+ year={2024},
136
+ volume={},
137
+ number={},
138
+ pages={10461-10465},
139
+ keywords={Systematics;Error analysis;Knowledge based systems;Oral communication;Signal processing;Data models;Acoustics;multi-talker;asr;synthetic conversation},
140
+ doi={10.1109/ICASSP48485.2024.10446589}}
141
+
142
+
143
+ ```
144
+
145
+ ### License
146
+
147
+ CC-BY-NC 4.0
148
+
149
+ ### Contact
150
+
151
+ Contributions are welcome; feel free to create a PR or email me:
152
+
153
+ ```
154
+ [Binh Nguyen](nguyenvulebinh[at]gmail.com)
155
+ ```