manasdhir04 commited on
Commit
4fc0e48
·
verified ·
1 Parent(s): c6b2005

Add public test split (151 clips) and dataset card

Browse files
Files changed (2) hide show
  1. README.md +141 -0
  2. data/test-00000-of-00001.parquet +3 -0
README.md ADDED
@@ -0,0 +1,141 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ dataset_info:
3
+ features:
4
+ - name: audio
5
+ dtype: audio
6
+ - name: id
7
+ dtype: string
8
+ - name: spoken_text
9
+ dtype: string
10
+ - name: intended_text
11
+ dtype: string
12
+ - name: domain
13
+ dtype: string
14
+ - name: audio_length_s
15
+ dtype: float64
16
+ - name: speaker_id
17
+ dtype: int64
18
+ - name: gender
19
+ dtype: string
20
+ - name: age
21
+ dtype: int64
22
+ - name: current_city
23
+ dtype: string
24
+ splits:
25
+ - name: test
26
+ num_examples: 151
27
+ configs:
28
+ - config_name: default
29
+ data_files:
30
+ - split: test
31
+ path: data/test-*
32
+ pretty_name: Voice Arena Dictation Dataset (public test)
33
+ language:
34
+ - en
35
+ license: cc-by-nc-4.0
36
+ task_categories:
37
+ - automatic-speech-recognition
38
+ size_categories:
39
+ - n<1K
40
+ tags:
41
+ - speech
42
+ - asr
43
+ - benchmark
44
+ - dictation
45
+ - english
46
+ - en-US
47
+ - inverse-text-normalization
48
+ - disfluency
49
+ ---
50
+
51
+ # Voice Arena Dictation Dataset (public test)
52
+
53
+ A US English **dictation** test set: people speaking the kind of text they would
54
+ otherwise type — emails, chat messages, notes, code requests and prompts for AI
55
+ assistants. Every clip carries two references:
56
+
57
+ - `spoken_text` — what was actually said, verbatim, including self-corrections
58
+ ("seven thousand, wait no, four thousand"), restarts and hedges.
59
+ - `intended_text` — what should land on the page: corrections applied, numbers,
60
+ dates, currency, emails and URLs written out in their final form.
61
+
62
+ The gap between the two is the point of the set. A system can transcribe every word
63
+ correctly and still produce text the speaker would have to fix by hand.
64
+
65
+ This is the **public** split, released for self-scoring. A speaker-disjoint private
66
+ split is withheld to limit benchmark-specific optimisation.
67
+
68
+ ## At a glance
69
+
70
+ | | |
71
+ | :--- | :--- |
72
+ | Clips | 151 |
73
+ | Duration | 0.73 h |
74
+ | Speakers | 71 |
75
+ | Clip length | 17.5 s mean / 12.0 s median (3.1 s – 90.3 s) |
76
+ | Words (spoken) | 6,922 |
77
+ | Female / male / other clips | 124 / 25 / 2 |
78
+ | Speaker age | 23 – 72 (median 47) |
79
+ | US states (current residence) | 28 |
80
+ | Clips where spoken ≠ intended | 121 |
81
+
82
+ ### Domains
83
+
84
+ | `domain` | Clips |
85
+ | :--- | ---: |
86
+ | `code` | 26 |
87
+ | `personal_messaging` | 25 |
88
+ | `work_chat` | 24 |
89
+ | `ai_prompt` | 22 |
90
+ | `docs_notes` | 22 |
91
+ | `email` | 22 |
92
+ | `long_form` | 10 |
93
+
94
+ ### Speech phenomena
95
+
96
+ | | Clips |
97
+ | :--- | ---: |
98
+ | Named entities (people, dates, amounts, emails/URLs, …) | 74 |
99
+ | Questions | 43 |
100
+ | Self-corrections | 27 |
101
+ | Spoken lists | 26 |
102
+ | Restarts | 23 |
103
+ | Hedges | 19 |
104
+
105
+ ## Loading
106
+
107
+ ```python
108
+ from datasets import load_dataset
109
+
110
+ ds = load_dataset("VoiceArena/Dictation_Dataset", split="test")
111
+ print(ds[0]["spoken_text"])
112
+ print(ds[0]["intended_text"])
113
+ ```
114
+
115
+ Audio is mono FLAC at 44.1 kHz.
116
+
117
+ ## Data fields
118
+
119
+ | Field | Type | Description |
120
+ | :--- | :--- | :--- |
121
+ | `audio` | Audio | The speech clip (FLAC). |
122
+ | `id` | string | Unique clip identifier. |
123
+ | `spoken_text` | string | Verbatim reference of what was said, including self-corrections, restarts and hedges. |
124
+ | `intended_text` | string | The text the speaker meant to produce: corrections resolved and written-form formatting applied. |
125
+ | `domain` | string | Writing context: `email`, `work_chat`, `personal_messaging`, `docs_notes`, `code`, `ai_prompt` or `long_form`. |
126
+ | `audio_length_s` | float | Clip duration in seconds. |
127
+ | `speaker_id` | int | Anonymised speaker identifier, consistent across all clips from the same speaker. |
128
+ | `gender` | string | Speaker's self-reported gender: `female`, `male` or `other`. |
129
+ | `age` | int | Speaker age in years. |
130
+ | `current_city` | string | Speaker's city of residence as "city, state, county". |
131
+
132
+ ## Scoring
133
+
134
+ Two ways to use the references:
135
+
136
+ - **Transcription** — compare the model output against `spoken_text` with word error rate.
137
+ - **Dictation** — compare against `intended_text`. This rewards systems that resolve
138
+ self-corrections and format numbers, dates and addresses the way a person would type them.
139
+
140
+ Report the **aggregate** over the split — `sum(ins + del + sub) / sum(reference words)` —
141
+ not the average of per-clip rates.
data/test-00000-of-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0b13f09bc5f5a326df0a7b002159b627ddcba37af0f19bbc413934c3a0bbfe80
3
+ size 105738127