gilkeyio commited on
Commit
0daa1eb
1 Parent(s): 757c9f7

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +36 -8
README.md CHANGED
@@ -94,9 +94,9 @@ Librispeech is a corpus of read English speech, designed for training and evalua
94
  The Montreal Forced Aligner (MFA) was used to generate word and phoneme level alignments for the Librispeech dataset.
95
 
96
 
97
- - **Curated by:** Vassil Panayotov, Guoguo Chen, Daniel Povey, Sanjeev Khudanpur
98
  - **Funded by:** DARPA LORELEI
99
- - **Shared by:** Loren Lugosch
100
  - **Language(s) (NLP):** English
101
  - **License:** Creative Commons Attribution 4.0 International License
102
 
@@ -118,10 +118,32 @@ The dataset only contains read speech, so may not perform as well on spontaneous
118
 
119
  ## Dataset Structure
120
 
121
- The dataset contains 1000 hours of segmented read English speech from audiobooks. There are three subsets: 100 hours (train-clean-100), 360 hours (train-clean-360) and 500 hours (train-other-500).
122
 
123
  The alignments connect the audio to the reference text transcripts on word and phoneme level.
124
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
125
  ## Dataset Creation
126
 
127
  ### Curation Rationale
@@ -143,10 +165,11 @@ The audiobooks are read by volunteers for the LibriVox project. Information abou
143
  #### Annotation process
144
 
145
  The Montreal Forced Aligner was used to create word and phoneme level alignments between the audio and reference texts. The aligner is based on Kaldi.
 
146
 
147
  #### Who are the annotators?
148
 
149
- The alignments were generated automatically by the Montreal Forced Aligner and shared by Loren Lugosch.
150
 
151
  #### Personal and Sensitive Information
152
 
@@ -158,11 +181,11 @@ The dataset contains only read speech from published books, not natural conversa
158
 
159
  ### Recommendations
160
 
161
- Users should understand that the alignments may contain errors and account for this in applications.
162
 
163
  ## Citation
164
 
165
- **BibTeX for Librispeech:**
166
  ```
167
  @inproceedings{panayotov2015librispeech,
168
  title={Librispeech: an ASR corpus based on public domain audio books},
@@ -173,7 +196,12 @@ Users should understand that the alignments may contain errors and account for t
173
  }
174
  ```
175
 
176
- **APA for Alignments:**
 
 
 
 
 
177
  ```
178
- Lugosch, L. (2019). LibriSpeech Alignments (1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.2619474
179
  ```
 
94
  The Montreal Forced Aligner (MFA) was used to generate word and phoneme level alignments for the Librispeech dataset.
95
 
96
 
97
+ - **Curated by:** Vassil Panayotov, Guoguo Chen, Daniel Povey, Sanjeev Khudanpur (for Librispeech)
98
  - **Funded by:** DARPA LORELEI
99
+ - **Shared by:** Loren Lugosch (for Alignments)
100
  - **Language(s) (NLP):** English
101
  - **License:** Creative Commons Attribution 4.0 International License
102
 
 
118
 
119
  ## Dataset Structure
120
 
121
+ The dataset contains 1000 hours of segmented read English speech from audiobooks. There are three train subsets: 100 hours (train-clean-100), 360 hours (train-clean-360) and 500 hours (train-other-500).
122
 
123
  The alignments connect the audio to the reference text transcripts on word and phoneme level.
124
 
125
+ ### Data Fields
126
+
127
+ - sex: M for male, F for female
128
+
129
+ - subset: dev_clean, dev_other, test_clean, test_other, train_clean_100, train_clean_360, train_other_500
130
+
131
+ - id: unique id of the data sample. (speaker id)-(chapter-id)-(utterance-id)
132
+
133
+ - audio: the audio, 16kHz
134
+
135
+ - transcript: the spoken text of the dataset, normalized and lowercased
136
+
137
+ - words: a list of words with fields:
138
+ - word: the text of the word
139
+ - start: the start time in seconds
140
+ - end: the end time in seconds
141
+
142
+ - phonemes: a list of phonemes with fields:
143
+ - phoneme: the phoneme spoken
144
+ - start: the start time in seconds
145
+ - end: the end time in seconds
146
+
147
  ## Dataset Creation
148
 
149
  ### Curation Rationale
 
165
  #### Annotation process
166
 
167
  The Montreal Forced Aligner was used to create word and phoneme level alignments between the audio and reference texts. The aligner is based on Kaldi.
168
+ In the process of formatting this into a HuggingFace dataset, words with empty text and phonemes with empty text, silence tokens, or spacing tokens were removed
169
 
170
  #### Who are the annotators?
171
 
172
+ The alignments were generated automatically by the Montreal Forced Aligner and shared by Loren Lugosch. The TextGrid files were parsed and integrated into this dataset by Kim Gilkey.
173
 
174
  #### Personal and Sensitive Information
175
 
 
181
 
182
  ### Recommendations
183
 
184
+ Users should understand that the alignments may contain errors and account for this in applications. For example, be wary of <UNK> tokens.
185
 
186
  ## Citation
187
 
188
+ **Librispeech:**
189
  ```
190
  @inproceedings{panayotov2015librispeech,
191
  title={Librispeech: an ASR corpus based on public domain audio books},
 
196
  }
197
  ```
198
 
199
+ **Librispeech Alignments:**
200
+ ```
201
+ Loren Lugosch, Mirco Ravanelli, Patrick Ignoto, Vikrant Singh Tomar, and Yoshua Bengio, "Speech Model Pre-training for End-to-End Spoken Language Understanding", Interspeech 2019.
202
+ ```
203
+
204
+ **Montreal Forced Aligner:**
205
  ```
206
+ Michael McAuliffe, Michaela Socolof, Sarah Mihuc, Michael Wagner, and Morgan Sonderegger. "Montreal Forced Aligner: trainable text-speech alignment using Kaldi", Interspeech 2017.
207
  ```