internetoftim commited on
Commit
5ad6943
·
verified ·
1 Parent(s): 2c98145

Update dataset card (pld.md)

Browse files
Files changed (1) hide show
  1. README.md +199 -69
README.md CHANGED
@@ -10,49 +10,119 @@ language:
10
  - pam
11
  - tsg
12
  - war
 
 
 
 
 
13
  task_categories:
14
  - automatic-speech-recognition
15
  - text-to-speech
 
 
16
  tags:
17
  - philippines
18
  - philippine-languages
19
  - low-resource
20
- - speech
21
  - multilingual
22
- size_categories:
23
- - 100K<n<1M
 
 
 
 
 
 
 
24
  ---
25
 
26
  # Philippine Language Dataset (PLD)
27
 
28
- Utterance-level recordings from the University of the Philippines Digital
29
- Signal Processing Laboratory's Philippine Language Dataset, packaged as a
30
- Hugging Face Parquet dataset with 16kHz mono audio.
 
 
 
 
 
 
31
 
32
- Every row is one prompted recording: the corpus ships pre-segmented WAVs with
33
- the prompt text stored inline in each session log, so no forced alignment or
34
- segmentation was applied. Text is normalized only for whitespace and unicode
35
- punctuation — orthography and dialectal spelling are preserved as recorded.
36
 
37
- **334,268 utterances · 448.2 hours · 980 speakers · 10 languages**
 
 
 
38
 
39
- ## ⚠️ Read before training: `text_is_prompt`
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
40
 
41
  Rows where `speech_type == "spontaneous"` do **not** carry a transcript. The
42
- session logs store the *elicitation question* that was put to the speaker
43
- (e.g. "Saen an dream destination mo?"), while the audio is 20-90 seconds of
44
  their free-speech answer. The same question text repeats verbatim across
45
- different speakers. These rows are flagged `text_is_prompt = true` and must be
46
- excluded from any supervised (audio, text) training:
 
 
 
47
 
48
  ```python
49
  ds = ds.filter(lambda x: not x["text_is_prompt"])
50
  ```
51
 
52
- They remain in the dataset because the audio is genuine spontaneous speech,
53
- useful for self-supervised pretraining, speaker modeling, or re-transcription.
54
 
55
- ## Languages
 
 
 
 
 
 
 
 
 
 
 
56
 
57
  | Code | Language | Utterances | Hours |
58
  |---|---|---|---|
@@ -67,53 +137,34 @@ useful for self-supervised pretraining, speaker modeling, or re-transcription.
67
  | `pag` | Pangasinan | 5,566 | 8.7 |
68
  | `tsg` | Tausug | 4,495 | 5.9 |
69
 
70
- English word and sentence lists (`EngW.txt`, `EngSen.txt`) were read by the
71
- same speakers. Those rows are labeled `language = "eng"`, with
72
- `corpus_language` retaining the Philippine language collection they came from,
73
- so per-language filters stay clean.
74
 
75
- ## Speech types
76
 
77
- | Type | Utterances | Hours | Mean duration |
78
- |---|---|---|---|
79
- | `read` | 158,121 | 302.9 | 6.9s |
80
- | `isolated` | 164,447 | 107.3 | 2.3s |
81
- | `spontaneous` | 2,586 | 32.2 | 44.8s |
82
- | `digits` | 9,114 | 5.9 | 2.3s |
83
-
84
- - `read` — full prompted sentences (news, medical, literature, essays); the
85
- material best suited to TTS
86
- - `isolated` — single words and short phrases from word lists
87
- - `digits` — spoken digit strings
88
- - `spontaneous` — free speech; **prompt-only text**, see the warning above
89
 
90
- ## Usage
91
 
92
- **ASR finetuning (Whisper):**
93
- ```python
94
- from datasets import load_dataset, Audio
95
- ds = load_dataset("sapinsapin/pld")
96
- ds = ds.filter(lambda x: not x["text_is_prompt"] and 0.3 <= x["duration"] <= 30.0)
97
- ds = ds.cast_column("audio", Audio(sampling_rate=16000))
98
- ```
99
-
100
- **TTS finetuning (read speech only):**
101
- ```python
102
- ds = ds.filter(lambda x: x["speech_type"] == "read"
103
- and 1.0 <= x["duration"] <= 15.0
104
- and x["num_words"] >= 3)
105
- ```
106
 
107
- **Single language:**
108
- ```python
109
- ds = ds.filter(lambda x: x["language"] == "bcl") # Bikol
110
- ```
111
 
112
  ## Schema
113
 
114
  | Field | Type | Description |
115
  |---|---|---|
116
- | `audio` | `Audio(16000)` | 16kHz mono WAV |
117
  | `sentence` | `str` | Prompt text read by the speaker (see `text_is_prompt`) |
118
  | `duration` | `float` | Seconds |
119
  | `num_words` | `int` | Whitespace word count |
@@ -128,27 +179,106 @@ ds = ds.filter(lambda x: x["language"] == "bcl") # Bikol
128
  | `gender` | `str` | `male` / `female` / `unknown` |
129
  | `age` | `int` | Speaker age, `-1` when unrecorded |
130
  | `speaker_dialect` | `str` | Self-reported dialect |
131
- | `mother_dialect` / `father_dialect` | `str` | Parents' dialects |
132
  | `profession` | `str` | Self-reported profession |
133
  | `session_id` | `str` | Recording session identifier |
134
  | `session_environment` | `str` | Recording environment note |
135
  | `source_file` | `str` | Original WAV stem |
136
 
137
- ## Splits
 
 
138
 
139
- | Split | Rows |
140
- |---|---|
141
- | `train` | 300,842 |
142
- | `test` | 33,426 |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
143
 
144
- Split is a seeded random 90/10 over utterances, so speakers appear in both
145
- splits. For speaker-disjoint evaluation, re-split on `speaker_id`.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
146
 
147
- ## Source
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
148
 
149
- Collected by the UP Diliman Digital Signal Processing Laboratory. Please cite
150
- the original corpus authors when using this data.
151
 
152
- ## Code
 
153
 
154
- Processing code: https://github.com/sapinsapin/pretraining
 
 
 
10
  - pam
11
  - tsg
12
  - war
13
+ license: other
14
+ license_name: up-dsp-research
15
+ pretty_name: Philippine Language Dataset (PLD)
16
+ size_categories:
17
+ - 100K<n<1M
18
  task_categories:
19
  - automatic-speech-recognition
20
  - text-to-speech
21
+ multilinguality:
22
+ - multilingual
23
  tags:
24
  - philippines
25
  - philippine-languages
26
  - low-resource
 
27
  - multilingual
28
+ - speech
29
+ - bikol
30
+ - cebuano
31
+ - kapampangan
32
+ - ilocano
33
+ - hiligaynon
34
+ - waray
35
+ - pangasinan
36
+ - tausug
37
  ---
38
 
39
  # Philippine Language Dataset (PLD)
40
 
41
+ **Ten Philippine languages, 980 speakers, 448 hours of prompted speech — one of the largest multilingual Philippine speech collections available as Parquet.**
42
+
43
+ <div align="center">
44
+
45
+ **334,268 utterances · 448.2 hours · 980 speakers · 10 languages · 16kHz mono**
46
+
47
+ [![Code](https://img.shields.io/badge/pipeline-github-black)](https://github.com/sapinsapin/halohalo)
48
+
49
+ </div>
50
 
51
+ Collected by the **University of the Philippines Diliman Digital Signal
52
+ Processing Laboratory**. Every row is one prompted recording: the corpus ships
53
+ pre-segmented WAVs with the prompt text stored inline in each session log, so
54
+ no forced alignment or segmentation was applied here.
55
 
56
+ Most Philippine language speech data stops at Tagalog. This one covers Bikol,
57
+ Cebuano, Kapampangan, Hiligaynon, Ilocano, Waray, Pangasinan and Tausug at
58
+ scale — languages with tens of millions of speakers and almost no public ASR/TTS
59
+ data.
60
 
61
+ > **Read [`text_is_prompt`](#-read-this-before-training) before you train.** 2,586 rows
62
+ > carry an elicitation *question* instead of a transcript, and silently training
63
+ > on them will poison your model.
64
+
65
+ ---
66
+
67
+ ## 30-second quickstart
68
+
69
+ ```python
70
+ from datasets import load_dataset
71
+
72
+ ds = load_dataset("sapinsapin/pld", split="train", streaming=True)
73
+ row = next(iter(ds))
74
+
75
+ print(row["language_name"], "|", row["sentence"])
76
+ print(row["speech_type"], row["duration"], "s")
77
+ ```
78
+
79
+ The training-ready filter, in full:
80
+
81
+ ```python
82
+ ds = load_dataset("sapinsapin/pld", split="train")
83
+ ds = ds.filter(lambda x: not x["text_is_prompt"] and 0.3 <= x["duration"] <= 30.0)
84
+ ```
85
+
86
+ One language at a time:
87
+
88
+ ```python
89
+ bikol = ds.filter(lambda x: x["language"] == "bcl")
90
+ ```
91
+
92
+ ---
93
+
94
+ ## ⚠️ Read this before training
95
+
96
+ ### `text_is_prompt` — 2,586 rows have no transcript
97
 
98
  Rows where `speech_type == "spontaneous"` do **not** carry a transcript. The
99
+ session logs store the *elicitation question* that was put to the speaker —
100
+ e.g. *"Saen an dream destination mo?"* — while the audio is 20–90 seconds of
101
  their free-speech answer. The same question text repeats verbatim across
102
+ different speakers.
103
+
104
+ They are kept because the audio is genuine spontaneous speech (32 hours of it,
105
+ valuable for pretraining, VAD, diarization, or re-transcription), but they are
106
+ poison for supervised `(audio, text)` training:
107
 
108
  ```python
109
  ds = ds.filter(lambda x: not x["text_is_prompt"])
110
  ```
111
 
112
+ ### Other things to know
 
113
 
114
+ | Gotcha | Detail | What to do |
115
+ |---|---|---|
116
+ | Prompts, not transcripts | Text is what the speaker was *asked* to read; no one verified they read it exactly | Treat as weakly-supervised; round-trip ASR to score |
117
+ | Speaker overlap | Random 90/10 split over utterances, so speakers appear in both splits | Re-split on `speaker_id` for speaker-disjoint eval |
118
+ | `eng` is not native English | English word/sentence lists read by Filipino L2 speakers | Use `corpus_language` to see which collection they came from |
119
+ | Half the corpus is single words | 164k `isolated` rows average 2.3 s | Filter on `speech_type` for sentence-level work |
120
+
121
+ ---
122
+
123
+ ## What's inside
124
+
125
+ ### Languages
126
 
127
  | Code | Language | Utterances | Hours |
128
  |---|---|---|---|
 
137
  | `pag` | Pangasinan | 5,566 | 8.7 |
138
  | `tsg` | Tausug | 4,495 | 5.9 |
139
 
140
+ English word and sentence lists (`EngW.txt`, `EngSen.txt`) were read by the same
141
+ speakers. Those rows are labeled `language = "eng"`, while `corpus_language`
142
+ retains the Philippine language collection they came from, so per-language
143
+ filters stay clean either way.
144
 
145
+ ### Speech types
146
 
147
+ | Type | Utterances | Hours | Mean | What it is |
148
+ |---|---|---|---|---|
149
+ | `read` | 158,121 | 302.9 | 6.9 s | Full prompted sentences — news, medical, literature, education, tourism. **Best material for TTS.** |
150
+ | `isolated` | 164,447 | 107.3 | 2.3 s | Single words and short phrases from word lists |
151
+ | `spontaneous` | 2,586 | 32.2 | 44.8 s | Free speech — **prompt-only text**, see the warning above |
152
+ | `digits` | 9,114 | 5.9 | 2.3 s | Spoken digit strings |
 
 
 
 
 
 
153
 
154
+ ### Splits
155
 
156
+ | Split | Rows |
157
+ |---|---|
158
+ | `train` | 300,842 |
159
+ | `test` | 33,426 |
 
 
 
 
 
 
 
 
 
 
160
 
161
+ ---
 
 
 
162
 
163
  ## Schema
164
 
165
  | Field | Type | Description |
166
  |---|---|---|
167
+ | `audio` | `Audio(16000)` | 16 kHz mono, FLAC-compressed in storage |
168
  | `sentence` | `str` | Prompt text read by the speaker (see `text_is_prompt`) |
169
  | `duration` | `float` | Seconds |
170
  | `num_words` | `int` | Whitespace word count |
 
179
  | `gender` | `str` | `male` / `female` / `unknown` |
180
  | `age` | `int` | Speaker age, `-1` when unrecorded |
181
  | `speaker_dialect` | `str` | Self-reported dialect |
182
+ | `mother_dialect` / `father_dialect` | `str` | Parents' dialects — useful for contact/variation studies |
183
  | `profession` | `str` | Self-reported profession |
184
  | `session_id` | `str` | Recording session identifier |
185
  | `session_environment` | `str` | Recording environment note |
186
  | `source_file` | `str` | Original WAV stem |
187
 
188
+ The dialect fields are unusually rich for a speech corpus — speaker, mother and
189
+ father dialect are all recorded, which supports dialectometry and language-contact
190
+ work that most corpora can't.
191
 
192
+ ---
193
+
194
+ ## Models trained on this data
195
+
196
+ No reference finetune has been trained on PLD yet. The pipeline supports it —
197
+ add a `pld` branch to the dataset adapter and the existing trainers work
198
+ unchanged.
199
+
200
+ Meanwhile, the sibling Filipino corpus has working baselines you can copy the
201
+ recipe from:
202
+
203
+ | Model | Task | Trained on |
204
+ |---|---|---|
205
+ | [`sapinsapin/speecht5_tts-fsc`](https://huggingface.co/sapinsapin/speecht5_tts-fsc) | Text-to-speech | [`filipinospeechcorpus`](https://huggingface.co/datasets/sapinsapin/filipinospeechcorpus) |
206
+ | [`sapinsapin/whisper-small-fsc`](https://huggingface.co/sapinsapin/whisper-small-fsc) | Speech recognition | [`filipinospeechcorpus`](https://huggingface.co/datasets/sapinsapin/filipinospeechcorpus) |
207
+
208
+ **If you train something on PLD, tag this dataset in your model card** — it will
209
+ show up here automatically, and a Bikol or Cebuano ASR baseline would be the
210
+ first of its kind in public.
211
 
212
+ ---
213
+
214
+ ## How it was built
215
+
216
+ 1. Walk every session directory; parse the per-session `.log` (speaker
217
+ demographics header, then one row per utterance: WAV name, prompt list,
218
+ prompt text).
219
+ 2. Classify each utterance's `speech_type`. Explicit markers (`_Iso_`, `_Utt_`,
220
+ `Spontaneous`, digits) are used where present; the corpus uses at least five
221
+ naming conventions, so the ~55k rows with no marker are typed by the
222
+ **measured median word count** of their prompt list rather than by guessing
223
+ from the filename.
224
+ 3. Repair double-encoded UTF-8 in transcripts (`hapúnan` → `hap��nan`) — 150 of
225
+ 166 affected lines recover; the rest are left intact rather than risk a worse
226
+ string.
227
+ 4. Resample to 16 kHz mono, encode FLAC, shard to Parquet, 90/10 random split.
228
+
229
+ 1,943 rows (0.6%) reference WAVs that are not present in the archive and were
230
+ skipped.
231
+
232
+ Pipeline source: [`process_pld_parquet.py`](https://github.com/sapinsapin/halohalo/blob/main/process_pld_parquet.py)
233
+ · parser: [`halolib/pld.py`](https://github.com/sapinsapin/halohalo/blob/main/halolib/pld.py)
234
+
235
+ ---
236
+
237
+ ## Limitations
238
+
239
+ - **Prompted, not conversational.** Except for the 32 h spontaneous portion,
240
+ this is people reading from lists. Prosody and vocabulary reflect that.
241
+ - **Transcripts are unverified prompts.** Nobody checked that speakers read the
242
+ prompt exactly; expect a residual mismatch rate.
243
+ - **Coverage is uneven** — Bikol has 95.8 h, Tausug 5.9 h. Don't expect balanced
244
+ multilingual behaviour without resampling.
245
+ - **Recording conditions vary** by session and are only loosely described in
246
+ `session_environment`.
247
+ - **No held-out speaker split** is provided by default.
248
+ - Language codes follow ISO 639-3; `fil` and `tgl` distinctions in the wild are
249
+ inconsistent, so filter on both if you merge with other corpora.
250
+
251
+ ---
252
 
253
+ ## Related datasets
254
+
255
+ Part of the **halohalo** Philippine-language speech family:
256
+
257
+ | Dataset | What it covers | Scale |
258
+ |---|---|---|
259
+ | **pld** *(this one)* | 10 Philippine languages, prompted | 334k utterances · 448 h |
260
+ | [`sapinsapin/filipinospeechcorpus`](https://huggingface.co/datasets/sapinsapin/filipinospeechcorpus) | Filipino studio read + spontaneous | 305k segments · 65 h |
261
+ | [`sapinsapin/halo-livestream`](https://huggingface.co/datasets/sapinsapin/halo-livestream) | Taglish code-switched livestream speech | seed release |
262
+
263
+ ---
264
+
265
+ ## License, source and citation
266
+
267
+ Collected by the **UP Diliman Digital Signal Processing Laboratory**. This is a
268
+ repackaging for research use; the underlying corpus terms are those of UP-DSP.
269
+ **Please credit the original collectors**, and contact UP-DSP for terms covering
270
+ uses beyond research.
271
+
272
+ If you represent UP-DSP and want attribution, terms, or access changed, please
273
+ open a discussion on this repo.
274
+
275
+ ---
276
 
277
+ ## Contributing
 
278
 
279
+ Eight of these ten languages have essentially no public ASR or TTS baseline.
280
+ That is the opportunity here.
281
 
282
+ - Train a baseline on any single language and tag this dataset in your model card
283
+ - Report bad rows via the **Community** tab (include `source_file` and `session_id`)
284
+ - Improve the pipeline: [github.com/sapinsapin/halohalo](https://github.com/sapinsapin/halohalo)