Upload croissant_rai_ViVoice34.json

croissant_rai_ViVoice34.json

@@ -0,0 +1,83 @@
+{
+  "@context": {
+    "@vocab": "https://schema.org/",
+    "@language": "en",
+    "sc": "https://schema.org/",
+    "rai": "http://mlcommons.org/croissant/RAI/",
+    "prov": "http://www.w3.org/ns/prov#"
+  },
+  "@type": "Dataset",
+  "name": "ViVoice34",
+  "description": "anonymous-vivoice34/ViVoice34 dataset hosted on Hugging Face and contributed by the HF Datasets community",
+  "alternateName": [
+    "anonymous-vivoice34/ViVoice34"
+  ],
+  "creator": {
+    "@type": "Person",
+    "name": "anonymous-vivoice34",
+    "url": "https://huggingface.co/anonymous-vivoice34"
+  },
+  "keywords": [
+    "apache-2.0",
+    "🇺🇸 Region: US"
+  ],
+  "license": "https://choosealicense.com/licenses/apache-2.0/",
+  "url": "https://huggingface.co/datasets/anonymous-vivoice34/ViVoice34",
+  "rai:dataLimitations": "ViVoice-34 has natural imbalances reflecting real-world data collection: Southern and Northern regions dominate (48.14h and 41.71h), while Central Vietnamese speech is underrepresented (18.90h). Children and older adults are significantly underrepresented compared to adults and middle-aged speakers. The dataset contains a wide range of recording conditions (from clean read speech to noisy real-world environments), but extremely low-quality or heavily overlapping multi-speaker recordings were filtered out. It is not recommended for medical speech applications, child speech modeling, or very elderly voice analysis due to limited coverage in those demographics. Aggregate metrics may mask regional performance gaps, especially for Central Vietnamese accents.",
+  "rai:dataBiases": "The dataset exhibits geographic imbalance, with Central Vietnamese speech being the smallest region. There is also gender-duration imbalance (male speech has longer total duration than female). Age distribution is skewed toward adults and adolescents; children and older adults are underrepresented. Data sources combine public web videos (which may favor more articulate or popular speakers) and real-world crowd-sourced recordings (which introduce device and environment variability). These biases may cause models to perform worse on underrepresented regions (especially Central accents), underrepresented age groups, and certain recording conditions.",
+  "rai:personalSensitiveInformation": "The dataset contains the following sensitive attributes at speaker level:\nGender, Age group, Geography (province of origin and region), Language (Vietnamese with regional dialects).\nNo health/medical data, political/religious beliefs, or direct personally identifiable information (names, faces, contact details) is included. All speaker identities are anonymized using hashed IDs. No raw personal information beyond the listed demographic and geographic metadata is released.",
+  "rai:dataUseCases": "ViVoice-34 is designed to measure and benchmark robustness of speaker verification, speaker identification, and speech recognition systems under real-world regional, demographic, and lexical variation in Vietnamese.\nValidated use cases:\n\nSpeaker verification & identification (speaker-disjoint, open-set)\nSpeaker attribute classification (gender, age group, dialect region)\nAutomatic Speech Recognition (ASR), especially robustness to regional words and loanwords\nDialect identification and region-aware modeling\n\nIt enables fine-grained analysis beyond aggregate metrics (region-stratified, lexical-stratified evaluation). It has not been validated for clinical/medical speech, child-directed speech, or extremely noisy environments.",
+  "rai:dataSocialImpact": "Positive impacts:\nViVoice-34 promotes inclusivity and fairness in Vietnamese speech technology by providing broad geographic coverage across all 34 provinces and rich demographic metadata. It helps developers build more robust systems that work well for diverse Vietnamese speakers, reducing bias against underrepresented accents (especially Central Vietnamese) and supporting equitable voice AI deployment in Vietnam.\nNegative risks & mitigations:\n\nPotential for misuse in surveillance or unauthorized speaker identification.\nRisk of amplifying existing regional stereotypes if models are not carefully evaluated.\nDemographic biases may lead to worse performance for children, elderly, or Central speakers if not addressed.\n\nMitigations: Public release under responsible license, detailed documentation of limitations/biases, encouragement of region-aware and fairness-aware evaluation. Dataset is anonymized.",
+  "rai:hasSyntheticData": true,
+  "prov:wasDerivedFrom": [
+    {
+      "@id": "https://huggingface.co/datasets/hustep-lab/VoxVietnam-Dataset",
+      "prov:label": "VoxVietnam",
+      "sc:license": "CC-BY-NC 4.0"
+    },
+    {
+      "@id": "https://huggingface.co/datasets/nguyendv02/ViMD_Dataset",
+      "prov:label": "ViMD (Multi-Dialect Vietnamese)",
+      "sc:license": "CC-BY-NC-ND 4.0"
+    },
+    {
+      "@id": "https://huggingface.co/datasets/hustep-lab/ViSEC",
+      "prov:label": "ViSEC",
+      "sc:license": "CC-BY 4.0"
+    }
+  ],
+  "prov:wasGeneratedBy": [
+    {
+      "@type": "prov:Activity",
+      "prov:type": {
+        "@id": "https://www.wikidata.org/wiki/Q4929239"
+      },
+      "prov:label": "Data Aggregation and Custom Recording",
+      "sc:description": "ViVoice-34 aggregates speech from four public Vietnamese datasets (VoxVietnam, ViMD, ViSEC, VietMed) and collects new real-world recordings from 450 participants via a custom web-based platform. Participants provided demographic metadata and recorded province-conditioned scripts containing regional vocabulary in natural environments using their own devices. Additionally, 500 synthetic speakers were generated using a commercial TTS API."
+    },
+    {
+      "@type": "prov:Activity",
+      "prov:type": {
+        "@id": "https://www.wikidata.org/wiki/Q5227332"
+      },
+      "prov:label": "Audio Quality and Speaker Cleaning",
+      "sc:description": "All audio underwent multi-stage preprocessing: quality filtering (remove low volume, high noise, corrupted files), single-speaker filtering using pyannote.audio diarization model, utterance duration standardization (remove <1s, merge short utterances to ~3s minimum), and volume normalization."
+    },
+    {
+      "@type": "prov:Activity",
+      "prov:type": {
+        "@id": "https://www.wikidata.org/wiki/Q5227332"
+      },
+      "prov:label": "Speaker Identity Consistency Cleaning",
+      "sc:description": "Similarity-based cleaning using a pretrained speaker verification encoder. Performed inter-speaker duplicate detection (merging/removing duplicate IDs) and intra-speaker outlier removal based on cosine similarity of embeddings to ensure reliable speaker labels."
+    },
+    {
+      "@type": "prov:Activity",
+      "prov:type": {
+        "@id": "https://www.wikidata.org/wiki/Q109719325"
+      },
+      "prov:label": "Speaker Attribute and Lexical Annotation",
+      "sc:description": "Speaker-level attributes (gender, age group, dialect region) were annotated using model-assisted approach: fine-tuned Wav2Vec2 classifiers generated preliminary labels, followed by human verification on low-confidence samples. Utterance-level lexical annotations (loanwords and local/regional words) were also added."
+    }
+  ]
+}