You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Please describe your research or application use case for this Hindi-English conversational turn-taking dataset.
Log in or Sign Up to review the conditions and access this dataset content.
Smart Turn Hinglish: Gold Human-Annotated Validation Benchmark (518 Clips)
This is the curated gold standard human validation benchmark for conversational turn-taking, endpointing, and mid-turn hesitation detection in Hindi-English (Hinglish) code-switched speech.
Dataset Summary
- Total Clips: 518 spontaneous conversational turns curated from the Hindi-English Conversational Audio Corpus (HiACC).
- Class Balance:
- Incomplete Turns (245 clips / 47.3%): Natural thinking pauses, mid-sentence conjunctions ("kyunki...", "agar...", "lekin...", "aur...", "matlab..."), and prosodic hesitations where the speaker has NOT finished speaking.
- Complete Turns (273 clips / 52.7%): Definitive sentence endpoints, completed answers, and turn closures.
- Audio Format: 16,000 Hz, Mono FLAC.
- Metadata:
metadata.jsonlcontaining:id: Unique turn ID (e.g.AD09006,CH03004).audio_path: Relative path to audio file (hin-eng/complete/...orhin-eng/incomplete/...).endpoint_bool: Boolean (truefor complete turn,falsefor incomplete hesitation).duration: Exact clip duration in seconds (1.0s to 8.0s).transcription: Spoken transcript in Devanagari / Latin script.
Access & Gating
This dataset is gated. Please fill out the access request form above with your organization and intended research use to request access.
- Downloads last month
- 10