Agri Diarization Segmenter (pyannote segmentation, fine-tuned for Indic farmer field audio)
A speaker-segmentation model for noisy, multi-speaker agricultural voice queries in Hindi, Telugu, and Odia.
It is pyannote/segmentation-3.0 fine-tuned on
human-annotated FarmerChat field recordings, and it is the one fine-tuned component of Digital Green's
model-agnostic voice pipeline for agricultural speech. It is used to isolate the farmer's own speech
(diarization → best-speaker selection) before an unmodified ASR model transcribes it.
Why fine-tune
No off-the-shelf diarizer transferred cleanly to open-air, low-end-phone farmer recordings across these three languages: for every stock system the dominant error is false alarm (hearing speech in background noise). Fine-tuning the segmenter on this audio collapses that false alarm.
Results (leakage-free, matched held-out split)
Scored on a frozen 411-clip test split the model never trained on, against the same human reference at a 0.25 s collar, alongside the stock and cloud diarizers on the same clips:
| Diarizer | DER @0.25 (lower is better) |
|---|---|
| This model (fine-tuned pyannote) | 0.214 |
| Azure (cloud) | 0.249 |
| pyannote 3.1 (stock) | 0.417 |
The improvement over stock is almost entirely a reduction in false alarm (0.364 → 0.118), at the cost of a small rise in missed speech (0.065 → 0.111). The validation-checkpoint DER was 0.1908 (epoch 82).
Intended use
- Speaker segmentation / diarization for agricultural voice pipelines, as the front end to target-speaker selection before ASR.
- CPU-inferable (~0.4 s per clip) — no GPU required at inference.
Training data
Fine-tuned on human-annotated FarmerChat field recordings (Hindi / Telugu / Odia), with speaker turns and timestamps placed by human annotators. Clip-level split: 1,884 train / 401 dev / 411 test. The raw training audio and speaker-turn files are not redistributed here because they are farmer field recordings (privacy); this repository ships the model weights and recipe only.
Limitations
- Adapted on Hindi, Telugu, and Odia field audio; transfer to unseen languages is not established.
- The false-alarm reduction trades a little recall on the rare clips with three or more speakers.
- The split is by clip, not demonstrated to be speaker- or session-disjoint.
Usage
from pyannote.audio import Model
from pyannote.audio.pipelines import SpeakerDiarization
# load the fine-tuned segmentation model from the local checkpoint
seg = Model.from_pretrained("seg-ft-epoch82-DER0.1908.ckpt")
pipeline = SpeakerDiarization(segmentation=seg, clustering="AgglomerativeClustering")
pipeline.instantiate({
"clustering": {"method": "centroid", "threshold": 0.7},
"segmentation": {"min_duration_off": 0.0},
})
diarization = pipeline("clip.wav")
for turn, _, speaker in diarization.itertracks(yield_label=True):
print(f"{turn.start:.2f}-{turn.end:.2f} {speaker}")
Tune the clustering threshold for your audio. The checkpoint is a pyannote/Lightning segmentation model.
The accompanying paper, Model-Agnostic and Language-Agnostic Voice Pipeline Improvement for the Agriculture Domain, is included in this repository (arXiv:2609.20504).
License and attribution
MIT, following the base model pyannote/segmentation-3.0
(Bredin, pyannote.audio). Please cite pyannote.audio and Digital Green's agricultural voice-pipeline paper
if you use this model.