ChristianYang commited on
Commit
651b5b8
·
verified ·
1 Parent(s): 38f4988

Card: add DDS source (194,400 rows / ~252h), CC-BY-NC-4.0, real clean-studio enhancement GT

Browse files
Files changed (1) hide show
  1. README.md +66 -45
README.md CHANGED
@@ -1,5 +1,5 @@
1
  ---
2
- license: cc-by-sa-4.0
3
  language:
4
  - en
5
  - zh
@@ -31,31 +31,36 @@ pairs four short **24 kHz mono FLAC** clips with aligned transcripts:
31
 
32
  - an **environment** sample (different speaker, same acoustic scene),
33
  - a **speaker** reference (same speaker as the target utterance),
34
- - a **speaker-enhanced** copy of the reference (MossFormer2 speech enhancement),
 
35
  - the target **speech** to synthesise,
36
 
37
  so a model can learn to generate an utterance with both a specified voice and a
38
  specified environment. This release supersedes the earlier
39
  [Env-TTS-SD-Corpus](https://huggingface.co/datasets/humanify/Env-TTS-SD-Corpus)
40
  with a richer schema, higher sample rate, per-clip ASR for all three contexts,
41
- and two additional meeting sources (AMI, AliMeeting).
 
42
 
43
  ## Dataset statistics
44
 
45
  | metric | value |
46
  | --- | ---: |
47
- | **rows** | 189,918 |
48
- | **on-disk size** | ~119 GB |
49
- | **Σ `speech_duration`** | **1,542,868.6 s** (**428.6 h**) |
50
- | **Σ (`environment_audio_duration` + `speaker_audio_duration` + `speech_duration`)** | **4,507,421.5 s** (**1,252.1 h**) |
51
 
52
- Duration totals were computed on 2026-05-27 by summing the float duration
53
- columns across all 239 parquet shards (no audio decode).
 
 
54
 
55
  ### Rows by source dataset
56
 
57
  | `dataset` | rows |
58
  | --- | ---: |
 
59
  | `m3sd` | 108,708 |
60
  | `aishell4` | 27,924 |
61
  | `alimeeting` | 25,087 |
@@ -67,29 +72,30 @@ columns across all 239 parquet shards (no audio decode).
67
 
68
  | column | type | description |
69
  | --- | --- | --- |
70
- | `environment_audio_source` | binary (FLAC 24 kHz mono) | acoustic-scene reference, 2.5–15 s, from a **different speaker** in the same session |
71
  | `environment_audio_duration` | float32 | seconds |
72
- | `environment_audio_text` | string | transcript of the environment clip (gold or Qwen3-ASR) |
73
  | `speaker_audio_source` | binary (FLAC 24 kHz mono) | speaker-identity reference, 2.5–15 s, **same speaker** as `speech` |
74
  | `speaker_audio_duration` | float32 | seconds |
75
  | `speaker_audio_text` | string | transcript of the speaker reference clip |
76
- | `speaker_audio_source_enhanced` | binary (FLAC 24 kHz mono) | MossFormer2-enhanced version of `speaker_audio_source` |
77
  | `text` | string | transcript of `speech` |
78
  | `speech` | binary (FLAC 24 kHz mono) | target utterance, 3–15 s |
79
  | `speech_duration` | float32 | seconds |
80
  | `language` | string | `zh` / `en` / `auto` |
81
- | `dataset` | string | `m3sd` / `aishell4` / `msdwild` / `chime6` / `ami` / `alimeeting` |
82
- | `conversation_id` | string | unique within the source dataset |
83
- | `speaker_id` | string | within-conversation diarisation label |
84
- | `env_id` | string | acoustic-scene identifier (usually `conversation_id`) |
85
  | `text_source` | string | `original`, `asr`, or `mixed` |
86
- | `asr_token_count` | int32 | Qwen3-ASR token count for `speech` |
87
  | `asr_mean_logprob` | float32 | mean log-prob per token for `speech` |
88
 
89
  ## Source corpora
90
 
91
- | dataset | nominal hours | language | transcripts in corpus |
92
  | --- | ---: | --- | --- |
 
93
  | **M3SD** (Wu et al., 2025) | 770 | zh / en mixed | ❌ → Qwen3-ASR |
94
  | **AISHELL-4** (Fu et al., 2021) | 120 | zh | ✅ TextGrid |
95
  | **MSDWILD** (Liu et al., 2022) | 80 | zh / en mixed | ❌ → Qwen3-ASR |
@@ -97,40 +103,48 @@ columns across all 239 parquet shards (no audio decode).
97
  | **AMI** (SDM, diarizers-community) | ~100 | en | ❌ → Qwen3-ASR |
98
  | **AliMeeting** (OpenSLR 119, far ch.0) | ~120 | zh | ✅ TextGrid |
99
 
 
 
 
 
 
 
 
 
 
100
  ## Processing pipeline
101
 
102
  Built with the streaming pipeline in
103
- [`env-tts-data-pipeline`](https://github.com/ChristianYang/env-tts-data-pipeline)
104
- (three parallel stages: **download → process → upload**):
 
105
 
106
  1. **download** — stream each source conversation (HF mirrors, OpenSLR tar
107
- streams, BaiduPCS for M3SD, etc.) into a bounded local cache; emit a JSON
108
- sentinel when audio is ready.
109
  2. **process** — resample to **24 kHz mono**, walk diarisation turns, emit
110
  3–15 s `speech` slices with a same-speaker reference (≥2.5 s) and a
111
- different-speaker environment slice (≥2.5 s, extended into surrounding audio
112
- when needed). Missing or split transcripts are re-labelled with
113
- **Qwen3-ASR-1.7B** (vLLM backend). Rows are written as snappy parquet shards
114
- (~800 rows / shard, 4 shards per HF commit group).
115
- 3. **upload** — `HfApi.upload_folder` per sealed group with rate limiting and
116
- resume-safe JSON state under `state/`.
117
- 4. **enhance** (second pass) — stream each published shard back, run
118
- **MossFormer2_SE_48K** on `speaker_audio_source`, and fill
119
- `speaker_audio_source_enhanced` at the same 24 kHz storage rate.
120
-
121
- DAIEN-style RIR+noise augmentation on the speaker reference is **disabled** in
122
- this release; environment/speaker decoupling relies on cross-speaker env sampling
123
- plus MossFormer2 enhancement instead.
124
 
125
  ## Licensing
126
 
127
- Released under **CC-BY-SA-4.0**, inheriting the most restrictive terms among
128
- sources. In particular:
129
 
 
 
130
  - **M3SD** — academic / non-commercial research only.
131
  - **MSDWILD** — X-LANCE research-only agreement.
132
  - AISHELL-4 (Apache-2.0), CHiME-6 (CC-BY-SA-4.0), AMI, and AliMeeting carry
133
- their respective open/research terms.
134
 
135
  Redistributing extracted audio requires complying with each upstream licence.
136
 
@@ -139,6 +153,13 @@ Redistributing extracted audio requires complying with each upstream licence.
139
  Please cite the source papers when using this corpus:
140
 
141
  ```bibtex
 
 
 
 
 
 
 
142
  @article{wu2025m3sd,
143
  title={M3SD: Multi-modal, Multi-scenario and Multi-language Speaker
144
  Diarization Dataset},
@@ -172,7 +193,8 @@ Please cite the source papers when using this corpus:
172
  ```
173
 
174
  ASR re-labelling uses [Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B).
175
- Speaker enhancement uses MossFormer2 (ClearVoice).
 
176
 
177
  ## Loading
178
 
@@ -181,21 +203,20 @@ from datasets import load_dataset
181
 
182
  ds = load_dataset("ChristianYang/Env-TTS-Clean", split="train", streaming=True)
183
  row = next(iter(ds))
184
- print(row["text"])
185
- print(row["speech_duration"])
186
  # Audio columns decode automatically when accessed (24 kHz mono).
 
 
 
187
  ```
188
 
189
  ## Files on disk
190
 
191
  ```
192
  data/
193
- group_00000/
194
  manifest.json
195
  data_000000.parquet
196
- data_000001.parquet
197
- ...
198
- group_00001/
199
  ...
200
  ```
201
 
 
1
  ---
2
+ license: cc-by-nc-4.0
3
  language:
4
  - en
5
  - zh
 
31
 
32
  - an **environment** sample (different speaker, same acoustic scene),
33
  - a **speaker** reference (same speaker as the target utterance),
34
+ - a **speaker-enhanced** copy of the reference (MossFormer2 enhancement — or, for
35
+ the **DDS** source, the **real clean-studio recording** of the speaker reference),
36
  - the target **speech** to synthesise,
37
 
38
  so a model can learn to generate an utterance with both a specified voice and a
39
  specified environment. This release supersedes the earlier
40
  [Env-TTS-SD-Corpus](https://huggingface.co/datasets/humanify/Env-TTS-SD-Corpus)
41
  with a richer schema, higher sample rate, per-clip ASR for all three contexts,
42
+ two meeting sources (AMI, AliMeeting), and a controlled read-speech source
43
+ (**DDS / DAPS**) that contributes real clean-studio enhancement ground truth.
44
 
45
  ## Dataset statistics
46
 
47
  | metric | value |
48
  | --- | ---: |
49
+ | **rows** | **384,318** (189,918 diarization + 194,400 DDS) |
50
+ | **on-disk size** | ~197 GB |
51
+ | **Σ `speech_duration`** | **~680 h** (428.6 h diarization + ~251.7 h DDS) |
52
+ | **Σ (`environment_audio_duration` + `speaker_audio_duration` + `speech_duration`)** | **~1,964 h** (1,252.1 h + ~712 h DDS) |
53
 
54
+ Diarization-source totals were computed exactly on 2026-05-27 (239 shards). The
55
+ **DDS** portion (194,400 rows, added 2026-06-09; train groups `group_00061`–
56
+ `group_00122`) is estimated from a sampled shard (mean per row: speech 4.66 s,
57
+ env 4.24 s, speaker 4.29 s).
58
 
59
  ### Rows by source dataset
60
 
61
  | `dataset` | rows |
62
  | --- | ---: |
63
+ | `dds` | 194,400 |
64
  | `m3sd` | 108,708 |
65
  | `aishell4` | 27,924 |
66
  | `alimeeting` | 25,087 |
 
72
 
73
  | column | type | description |
74
  | --- | --- | --- |
75
+ | `environment_audio_source` | binary (FLAC 24 kHz mono) | acoustic-scene reference, 2.5–15 s, from a **different speaker** in the same scene |
76
  | `environment_audio_duration` | float32 | seconds |
77
+ | `environment_audio_text` | string | transcript of the environment clip (gold / Qwen3-ASR / DAPS script) |
78
  | `speaker_audio_source` | binary (FLAC 24 kHz mono) | speaker-identity reference, 2.5–15 s, **same speaker** as `speech` |
79
  | `speaker_audio_duration` | float32 | seconds |
80
  | `speaker_audio_text` | string | transcript of the speaker reference clip |
81
+ | `speaker_audio_source_enhanced` | binary (FLAC 24 kHz mono) | de-environment'd speaker reference: **MossFormer2-enhanced** for diarization sources; the **real clean-studio recording** for DDS |
82
  | `text` | string | transcript of `speech` |
83
  | `speech` | binary (FLAC 24 kHz mono) | target utterance, 3–15 s |
84
  | `speech_duration` | float32 | seconds |
85
  | `language` | string | `zh` / `en` / `auto` |
86
+ | `dataset` | string | `dds` / `m3sd` / `aishell4` / `msdwild` / `chime6` / `ami` / `alimeeting` |
87
+ | `conversation_id` | string | unique within the source dataset (for DDS: `{room}__{device}__{channel}`) |
88
+ | `speaker_id` | string | within-scene speaker label (for DDS: DAPS speaker, e.g. `f1`, `m8`) |
89
+ | `env_id` | string | acoustic-scene identifier (for DDS: `dds__{room}__{device}__{channel}`) |
90
  | `text_source` | string | `original`, `asr`, or `mixed` |
91
+ | `asr_token_count` | int32 | Qwen3-ASR token count for `speech` (0 when `text_source=original`) |
92
  | `asr_mean_logprob` | float32 | mean log-prob per token for `speech` |
93
 
94
  ## Source corpora
95
 
96
+ | dataset | hours (≈, this release) | language | transcripts |
97
  | --- | ---: | --- | --- |
98
+ | **DDS — Device-Degraded Speech, DAPS portion** (Li & Yamagishi, 2021) | ~252 | en | ✅ DAPS scripts |
99
  | **M3SD** (Wu et al., 2025) | 770 | zh / en mixed | ❌ → Qwen3-ASR |
100
  | **AISHELL-4** (Fu et al., 2021) | 120 | zh | ✅ TextGrid |
101
  | **MSDWILD** (Liu et al., 2022) | 80 | zh / en mixed | ❌ → Qwen3-ASR |
 
103
  | **AMI** (SDM, diarizers-community) | ~100 | en | ❌ → Qwen3-ASR |
104
  | **AliMeeting** (OpenSLR 119, far ch.0) | ~120 | zh | ✅ TextGrid |
105
 
106
+ **DDS** is single-speaker **read speech** (not a diarization corpus): 20 DAPS
107
+ speakers re-recorded across **9 rooms × 3 microphones × 6 positions** (162
108
+ acoustic conditions). For each condition the target `speech` and the
109
+ `environment_audio_source` (a *different* speaker, same room/mic/position) are the
110
+ **device-degraded** recordings, while `speaker_audio_source_enhanced` is the
111
+ matching **clean-studio** recording — a real enhancement ground truth rather than
112
+ a MossFormer2 estimate. All DDS text comes from the DAPS scripts
113
+ (`text_source = original`).
114
+
115
  ## Processing pipeline
116
 
117
  Built with the streaming pipeline in
118
+ [`env-tts-data-pipeline`](https://github.com/ChristianYang/env-tts-data-pipeline).
119
+
120
+ Diarization sources — three parallel stages **download → process → upload**:
121
 
122
  1. **download** — stream each source conversation (HF mirrors, OpenSLR tar
123
+ streams, etc.) into a bounded local cache; emit a JSON sentinel when ready.
 
124
  2. **process** — resample to **24 kHz mono**, walk diarisation turns, emit
125
  3–15 s `speech` slices with a same-speaker reference (≥2.5 s) and a
126
+ different-speaker environment slice (≥2.5 s). Missing/split transcripts are
127
+ re-labelled with **Qwen3-ASR-1.7B**. Snappy parquet shards (~800 rows / shard,
128
+ 4 shards per HF commit group).
129
+ 3. **upload** — `HfApi.upload_folder` per sealed group, resume-safe.
130
+ 4. **enhance** (second pass) — **MossFormer2_SE_48K** on `speaker_audio_source`.
131
+
132
+ **DDS** uses a dedicated parallel-channel pass (`process-dds`): each (room,
133
+ device, position) condition is one acoustic scene; rows are assembled directly
134
+ from the parallel clean/degraded recordings, and `speaker_audio_source_enhanced`
135
+ is filled in-place with the real clean-studio clip (no second-pass MossFormer2).
 
 
 
136
 
137
  ## Licensing
138
 
139
+ Released under **CC-BY-NC-4.0** (non-commercial), inheriting the most restrictive
140
+ terms among sources. In particular:
141
 
142
+ - **DDS / DAPS** — CC-BY-NC-4.0 (non-commercial). *This is the binding term for
143
+ the whole release.*
144
  - **M3SD** — academic / non-commercial research only.
145
  - **MSDWILD** — X-LANCE research-only agreement.
146
  - AISHELL-4 (Apache-2.0), CHiME-6 (CC-BY-SA-4.0), AMI, and AliMeeting carry
147
+ their respective open / research terms.
148
 
149
  Redistributing extracted audio requires complying with each upstream licence.
150
 
 
153
  Please cite the source papers when using this corpus:
154
 
155
  ```bibtex
156
+ @article{li2021dds,
157
+ title={DDS: A new device-degraded speech dataset for speech enhancement},
158
+ author={Li, Haoyu and Yamagishi, Junichi},
159
+ journal={arXiv preprint arXiv:2109.07931},
160
+ year={2021}
161
+ }
162
+
163
  @article{wu2025m3sd,
164
  title={M3SD: Multi-modal, Multi-scenario and Multi-language Speaker
165
  Diarization Dataset},
 
193
  ```
194
 
195
  ASR re-labelling uses [Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B).
196
+ Speaker enhancement uses MossFormer2 (ClearVoice). DDS is built on the
197
+ [DAPS](https://zenodo.org/records/4660670) dataset (Mysore, 2015).
198
 
199
  ## Loading
200
 
 
203
 
204
  ds = load_dataset("ChristianYang/Env-TTS-Clean", split="train", streaming=True)
205
  row = next(iter(ds))
206
+ print(row["text"], row["dataset"], row["speech_duration"])
 
207
  # Audio columns decode automatically when accessed (24 kHz mono).
208
+
209
+ # Filter to a single source (e.g. the DDS read-speech rows):
210
+ dds = ds.filter(lambda r: r["dataset"] == "dds")
211
  ```
212
 
213
  ## Files on disk
214
 
215
  ```
216
  data/
217
+ group_00000/ ... group_00122/ # group_00061–00122 are DDS
218
  manifest.json
219
  data_000000.parquet
 
 
 
220
  ...
221
  ```
222