Correct provenance labels for recorded audio corpora

Author: dylanuys

src/gasbench/dataset/configs/real_audio.yaml

@@ -355,18 +355,55 @@ datasets:
   # -- sounds --
 
   - name: fakesound2-real
-    path: 34data/fakesound2-real
+    path: ZeyuXie/FakeSound2
     modality: audio
     media_type: real
     source_format: zip
     media_per_archive: 500
-    archives_per_dataset: 2
+    archives_per_dataset: 1
+    include_paths: [test/audio.zip]
+    archive_include_paths: [audio/genuine/]
     notes: FakeSound 2 — genuine environmental/ambient audio
     generator_family: real
     content_category: sounds
 
+  - name: fakesound2-recorded-edits
+    path: ZeyuXie/FakeSound2
+    modality: audio
+    media_type: real
+    source_format: zip
+    media_per_archive: 500
+    archives_per_dataset: 1
+    include_paths: [test/audio.zip]
+    archive_include_paths: [audio/addition/, audio/splicing/, audio/separation_lassnet/]
+    notes: FakeSound 2 — recorded additions, splices, and LASSNet mask-based separation without generative synthesis
+    generator_family: real
+    content_category: sounds
+
   # -- speech --
 
+  - name: thorsten-voice
+    path: Thorsten-Voice/TV-44kHz-Full
+    modality: audio
+    media_type: real
+    source_format: parquet
+    media_per_archive: 100
+    archives_per_dataset: 10
+    notes: Human German speech recordings by Thorsten Müller; training corpus, not synthesized output
+    generator_family: real
+    content_category: speech
+
+  - name: asvspoof2021-pa-eval
+    path: 34data/asvspoof2021-pa-eval
+    modality: audio
+    media_type: real
+    source_format: zip
+    media_per_archive: 500
+    archives_per_dataset: 2
+    notes: Human speech and physical replay recordings; replay does not synthesize speech
+    generator_family: real
+    content_category: diverse
+
   - name: MediaSpeech
     path: ymoslem/MediaSpeech
     modality: audio

src/gasbench/dataset/configs/synthetic_audio.yaml

@@ -50,16 +50,6 @@ datasets:
     generator_family: mixed-generation
     content_category: diverse
 
-  - name: asvspoof2021-pa-eval
-    path: 34data/asvspoof2021-pa-eval
-    modality: audio
-    media_type: synthetic
-    source_format: zip
-    media_per_archive: 500
-    archives_per_dataset: 2
-    notes: ASVspoof 2021 Physical Access track evaluation set
-    generator_family: mixed-generation
-    content_category: diverse
 
   - name: cfad-fake
     path: 34data/cfad-fake
@@ -359,13 +349,15 @@ datasets:
   # -- sounds --
 
   - name: fakesound2-fake
-    path: 34data/fakesound2-fake
+    path: ZeyuXie/FakeSound2
     modality: audio
     media_type: synthetic
     source_format: zip
     media_per_archive: 500
-    archives_per_dataset: 2
-    notes: FakeSound 2 — AI-generated environmental/ambient audio
+    archives_per_dataset: 1
+    include_paths: [test/audio.zip]
+    archive_include_paths: [audio/generation_, audio/inpainting_, audio/editing_, audio/separation_flowsep/]
+    notes: FakeSound 2 — generated, inpainted, generatively edited, and FlowSep-resynthesized audio
     generator_family: mixed-generation
     content_category: sounds
 
@@ -449,16 +441,6 @@ datasets:
     generator_family: mixed-generation
     content_category: speech
 
-  - name: thorsten-voice
-    path: Thorsten-Voice/TV-44kHz-Full
-    modality: audio
-    media_type: synthetic
-    source_format: parquet
-    media_per_archive: 100
-    archives_per_dataset: 10
-    notes: German TTS voice - Thorsten
-    generator_family: mixed-generation
-    content_category: speech
 
   - name: ultravoice-cjk
     path: tutu0604/UltraVoice