File size: 8,867 Bytes
24cec02
 
 
 
 
 
 
 
18846f8
24cec02
2c01fed
24cec02
92c1c95
3ffe5f5
8e058cc
 
92c1c95
 
 
 
 
 
 
 
 
24cec02
 
 
92c1c95
 
 
 
 
 
 
 
 
 
 
 
 
 
fbcc88a
 
54904a1
aacc2e2
 
 
 
 
6de8925
aacc2e2
 
0d502af
34fae80
 
 
 
 
0d502af
34fae80
aacc2e2
34fae80
 
 
 
 
 
 
 
aacc2e2
770701f
 
 
 
 
34fae80
770701f
6de8925
fbcc88a
1ad118d
 
 
 
 
 
6de8925
 
1ad118d
 
b909e1a
1ad118d
 
 
 
b909e1a
1ad118d
be65551
1ad118d
6de8925
 
 
 
 
556f7a1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6de8925
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4032e19
 
 
 
 
 
 
 
 
 
 
1ad118d
 
 
4032e19
 
1ad118d
 
 
 
 
92c1c95
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
---
task_categories:
- automatic-speech-recognition
language:
- ms
- en
- zh
- ta
- id
---
# Malaysian STT Whisper format

Up to 15k hours annotated, we done heavy postprocessing and post-translation to improve pseudolabeled Whisper Large V3.

Also include word level timestamp.

## Postprocessing

1. Check repetitive trigrams.
2. Verify Voice Activity using Silero-VAD.
3. Verify scores using Force Alignment.

## Post-translation

We use [mesolitica/nanot5-base-malaysian-translation-v2.1](https://huggingface.co/mesolitica/nanot5-base-malaysian-translation-v2.1).

## Dataset involved

1. [Malaysian context v1](https://huggingface.co/datasets/mesolitica/pseudolabel-malaya-speech-stt-train-whisper-large-v3-timestamp)
2. [Malaysian context v2](https://huggingface.co/datasets/mesolitica/pseudolabel-malaysian-youtube-whisper-large-v3-timestamp)
3. [Malay audiobook](https://huggingface.co/datasets/mesolitica/pseudolabel-nusantara-large-v3-timestamp)
4. [Singaporean context](https://huggingface.co/datasets/mesolitica/pseudolabel-imda-large-v3-timestamp)
5. [Indonesian context](https://huggingface.co/datasets/mesolitica/pseudolabel-indonesian-large-v3-timestamp)
6. [Mandarin audio](https://huggingface.co/datasets/mesolitica/pseudolabel-mandarin-large-v3-timestamp)
7. [Tamil audio](https://huggingface.co/datasets/mesolitica/pseudolabel-tamil-large-v3-timestamp)
8. [Science context](https://huggingface.co/datasets/mesolitica/pseudolabel-science-large-v3-timestamp)
9. [Malay sarawak](https://huggingface.co/datasets/malaysia-ai/sarawakmalay-whisper-format)
10. [Scripted Malay Daily Use Speech Corpus](https://huggingface.co/datasets/malaysia-ai/scripted-malay-daily-use-speech-corpus-whisper-format)
11. [Malay Conversational Speech Corpus](https://huggingface.co/datasets/malaysia-ai/malay-conversational-speech-corpus-whisper-format)
12. [Iban](https://huggingface.co/datasets/malaysia-ai/iban-whisper-format)
13. [Malay dialects](https://huggingface.co/datasets/mesolitica/pseudolabel-malay-dialects-large-v3-timestamp)

### Word level timestamp

1. Malaysian context v1, 658.54 hours.

```
{"audio_filename": "prepared-pseudolabel-malaya-chunks/2-0.mp3", "new_text": "<|startoftranscript|><|ms|><|transcribeprecise|><|0.00|> luar<|0.34|><|0.42|> kan<|0.60|><|1.78|> sebab<|2.06|><|2.24|> benda<|2.42|><|2.52|> ni<|2.58|><|2.70|> berlaku<|3.08|><|3.20|> contoh<|3.50|><|3.56|> kita<|3.72|><|3.80|> pergi<|3.98|><|4.10|> ke<|4.16|><|4.40|> ATM<|4.76|><|5.70|> yang<|5.80|><|5.84|> bukan<|6.02|><|6.10|> Islam,<|6.34|><|6.96|> siang<|7.12|><|7.18|> hari<|7.42|><|endoftext|>"}
```

- [Label](pseudolabel-malaya-whisper-word-timestamp.jsonl)
- [Audio](prepared-pseudolabel-malaya-chunks.zip)
  
2. Malaysian context v2, 8058.17 hours.

```
{"audio_filename": "prepared-pseudolabel-chunks/0-0.mp3", "new_text": "<|startoftranscript|><|ms|><|transcribeprecise|><|0.00|> tu<|0.04|><|0.20|> So<|0.26|><|0.70|> gaji<|0.96|><|1.06|> berbeza<|1.42|><|2.46|> Gaji<|2.86|><|endoftext|>"}
```

- [Label](pseudolabel-whisper-word-timestamp.jsonl)
- [Audio](prepared-pseudolabel_alignment)
  
3. Singaporean context, 1829.21 hours.

```
{"audio_filename": "prepared-imda-chunks/0-0.mp3", "new_text": "<|startoftranscript|><|en|><|transcribeprecise|><|0.00|> Households<|0.58|><|0.64|> with<|0.76|><|0.86|> target<|1.16|><|1.24|> sets<|1.50|><|1.70|> were<|1.82|><|1.90|> encouraged<|2.40|><|2.44|> to<|2.48|><|2.62|> try<|2.80|><|2.90|> keeping<|3.24|><|endoftext|>"}
```

- [Label](imda-whisper-word-timestamp.jsonl)
- [Audio](prepared-imda_alignment)
  
4. Science context, 4992.42 hours.

```
{"audio_filename": "prepared-science-chunks/0-0.mp3", "new_text": "<|startoftranscript|><|en|><|transcribeprecise|><|0.00|> Visual<|0.24|><|0.30|> Studio<|0.60|><|0.76|> Code<|1.00|><|1.06|> integration.<|1.68|><|3.46|> Here's<|3.70|><|3.76|> what<|3.88|><|3.94|> will<|4.06|><|4.10|> be<|4.14|><|4.28|> new.<|4.44|><|5.36|> You<|5.42|><|5.46|> will<|5.58|><|5.62|> have<|5.74|><|5.82|> more<|5.96|><|6.08|> choice<|6.40|><|6.48|> on<|6.52|><|6.60|> runtime<|6.96|><|7.02|> experiences.<|7.82|><|8.78|> Java<|9.06|><|9.16|> interoperability<|10.04|><|10.22|> will<|10.32|><|10.36|> be<|10.40|><|10.48|> available<|10.90|><|10.96|> on<|11.00|><|11.08|> all<|11.14|><|11.22|> platforms.<|11.78|><|12.22|> Objective<|12.68|><|12.78|> C<|12.78|><|13.00|> and<|13.06|><|13.16|> Swift<|13.42|><|13.48|> interoperability<|14.38|><|14.92|> will<|15.04|><|15.08|> be<|15.12|><|15.26|> supported<|15.70|><|15.78|> on<|15.82|><|15.90|> multiple<|16.26|><|16.34|> operating<|16.78|><|16.86|> systems.<|17.28|><|18.28|> Core<|18.62|><|18.74|> FX<|18.98|><|19.20|> will<|19.30|><|19.34|> be<|19.38|><|19.46|> extended<|19.96|><|20.30|> to<|20.34|><|20.42|> support<|20.74|><|20.84|> static<|21.16|><|21.22|> compilation<|22.04|><|endoftext|>"}
```
   
- [Label](prepared-science-word-timestamp.jsonl)
- [Audio](#)

## how to prepare the dataset

```bash
wget https://www.7-zip.org/a/7z2301-linux-x64.tar.xz
tar -xf 7z2301-linux-x64.tar.xz

# Malaysian context
wget https://huggingface.co/datasets/mesolitica/Malaysian-STT-Whisper/resolve/main/malaysian-stt.jsonl
huggingface-cli download --repo-type dataset \
--include 'output-audio-malaya.z*' \
--local-dir './' \
mesolitica/pseudolabel-malaya-speech-stt-train-whisper-large-v3-timestamp
./7zz x output-audio-malaya.zip -y -mmt40
huggingface-cli download --repo-type dataset \
--include 'output-audio.z*' \
--local-dir './' \
mesolitica/pseudolabel-malaysian-youtube-whisper-large-v3-timestamp
./7zz x output-audio.zip -y -mmt40

# Malay audiobook
wget https://huggingface.co/datasets/mesolitica/pseudolabel-nusantara-large-v3-timestamp/resolve/main/split-nusantara.zip
wget https://huggingface.co/datasets/mesolitica/pseudolabel-nusantara-large-v3-timestamp/resolve/main/prepared-nusantara.jsonl
unzip split-nusantara.zip

# Singaporean context
wget https://huggingface.co/datasets/mesolitica/pseudolabel-imda-large-v3-timestamp/resolve/main/prepared-imda.jsonl
wget https://huggingface.co/datasets/mesolitica/pseudolabel-imda-large-v3-timestamp/resolve/main/prepared-imda-ms.jsonl
huggingface-cli download --repo-type dataset \
--include '*.7z*' \
--local-dir './' \
mesolitica/IMDA-STT
./7zz x part1-mp3.7z.001 -y -mmt40
./7zz x part2-mp3.7z.001 -y -mmt40
./7zz x part3-same-audio-mp3.7z.001 -y -mmt40
./7zz x part3-separate-audio-mp3.7z.001 -y -mmt40
./7zz x part4-same-audio-mp3.7z.001 -y -mmt40
./7zz x part4-separate-audio-mp3.7z.001 -y -mmt40
./7zz x part5-same-audio-mp3.7z.001 -y -mmt40
./7zz x part5-separate-audio-mp3.7z.001 -y -mmt40
./7zz x part6-1-audio-mp3.7z.001 -y -mmt40
./7zz x part6-2-audio-mp3.7z.001 -y -mmt40
./7zz x part6-3-audio-mp3.7z.001 -y -mmt40

# Indonesian context
huggingface-cli download --repo-type dataset \
--include 'split-indonesian.z*' \
--local-dir './' \
mesolitica/pseudolabel-indonesian-large-v3-timestamp
./7zz x split-indonesian.zip -y -mmt40

# Science context
wget https://huggingface.co/datasets/mesolitica/Malaysian-STT-Whisper/resolve/main/science-en-stt.jsonl
wget https://huggingface.co/datasets/mesolitica/Malaysian-STT-Whisper/resolve/main/science-ms-stt.jsonl
huggingface-cli download --repo-type dataset \
--include 'audio-chunk.z*' \
--local-dir './' \
mesolitica/pseudolabel-science-large-v3-timestamp
./7zz x audio-chunk.zip -y -mmt40

# Malay sarawak
wget https://huggingface.co/datasets/malaysia-ai/sarawakmalay-whisper-format/resolve/main/sarawakmalay.zip
wget https://huggingface.co/datasets/malaysia-ai/sarawakmalay-whisper-format/resolve/main/dataset.json -O sarawakmalay.json
unzip sarawakmalay.zip

# for Scripted Malay Daily Use Speech Corpus
wget https://huggingface.co/datasets/malaysia-ai/scripted-malay-daily-use-speech-corpus-whisper-format/resolve/main/scripted-malay-daily-use-speech-corpus-whisper-format.zip
wget https://huggingface.co/datasets/malaysia-ai/scripted-malay-daily-use-speech-corpus-whisper-format/resolve/main/scripted-malay-daily-use-speech-corpus-whisper-format.json
unzip scripted-malay-daily-use-speech-corpus-whisper-format.zip

# Malay Conversational Speech Corpus
wget https://huggingface.co/datasets/malaysia-ai/malay-conversational-speech-corpus-whisper-format/resolve/main/malay-conversational-speech-corpus-whisper-format.zip
wget https://huggingface.co/datasets/malaysia-ai/malay-conversational-speech-corpus-whisper-format/resolve/main/malay-conversational-speech-corpus-whisper-format.json
unzip malay-conversational-speech-corpus-whisper-format.zip

# Iban
wget https://huggingface.co/datasets/malaysia-ai/iban-whisper-format/resolve/main/iban-wav.zip
wget https://huggingface.co/datasets/malaysia-ai/iban-whisper-format/resolve/main/iban-dataset.json
unzip iban-wav.zip
```

## Source code

Source code at https://github.com/mesolitica/malaysian-dataset/tree/master/speech-to-text-semisupervised/distilled-malaysian-whisper