Datasetpublic

DigiGreen/agri-voice-eval

Agricultural Voice Evaluation Set — Hindi, Telugu, Odia Human quality-checked FarmerChat field recordings with per-clip audio and human reference transcripts, used to evaluate Digital Green's agricultural voice pipeline stage by stage. Companion to the larger Agri STT Benchmarking Dataset, focused on the multi-speaker / noisy conditions that motivate speaker selection and enhancement. Each clip carries two references: the farmer's own (main-speaker) transcript — the… See the full description on the dataset page: https://huggingface.co/datasets/DigiGreen/agri-voice-eval.

sourceHugging Facecc-by-4.0updated 3d agoView on Hugging Face

0likes180downloads

Dataset Card

Agricultural Voice Evaluation Set — Hindi, Telugu, Odia

Human quality-checked FarmerChat field recordings with per-clip audio and human reference transcripts, used to evaluate Digital Green's agricultural voice pipeline stage by stage. Companion to the larger Agri STT Benchmarking Dataset, focused on the multi-speaker / noisy conditions that motivate speaker selection and enhancement.

Each clip carries two references: the farmer's own (main-speaker) transcript — the production-faithful scoring target — and the full per-speaker human transcription.

Format

Hugging Face audiofolder: audio/<uuid>.ogg audio files plus metadata.csv.

Column	Description
`file_name`	Path to the clip's audio (`audio/<uuid>.ogg`)
`uuid`	Clip identifier
`language`	hindi / telugu / odia
`n_speakers`	Human-annotated speaker count
`multi_speaker`	True if more than one speaker
`audio_issues`	Human-annotated conditions (background noise, overlap, cross-talk, …)
`main_speaker_label`	Which annotated speaker is the farmer (the query-bearer)
`reference_farmer`	Primary reference — the farmer main-speaker transcript
`reference_full`	Full human transcription of every speaker, with turn timestamps
`duration_sec`, `snr_estimate_db`	Basic acoustic metadata
`scored_in_paper`	True for the scored evaluation subset

python

from datasets import load_dataset
ds = load_dataset("DigiGreen/<repo-name>")  # audiofolder: audio + metadata

Curation

Each clip was quality-checked under a written protocol: audio-issue tagging, speaker count, a timestamped per-speaker transcription, identification of the main (farmer) speaker, and extraction of agricultural terms. The farmer main-speaker transcript is the scoring target because the service must answer the farmer's question — a recognizer that also transcribes bystanders is penalized against it.

Privacy and provenance

Audio is farmer field recordings, released for research following the same basis as the companion benchmark. Production storage locations and internal quality-control notes are not included, and transcripts were screened for obvious personal identifiers (phone numbers, self-identifying name phrases); none were found. If you have a concern about a specific clip, contact Digital Green for removal.

The accompanying paper, Model-Agnostic and Language-Agnostic Voice Pipeline Improvement for the Agriculture Domain, is included in this repository (arXiv:2609.20504).

License

CC-BY-4.0. Please cite Digital Green's agricultural voice-pipeline paper and the companion benchmark.