pyannote-seg-afr-childspeech-ft

Speaker diarization pipeline finetuned for Afrikaans child speech.

The segmentation model is a finetune of pyannote/segmentation-3.0 (PyanNet, powerset with max 2 concurrent speakers, 16 kHz, 15 s training chunks). Embedding and clustering follow the pyannote 3.1 recipe, run under pyannote-audio 4.0.4.

Requirements

• pyannote.audio == 4.0.4 (pinned — see Version notes below)
• A Hugging Face token, having accepted the terms for:
  • pyannote/speaker-diarization-community-1 — see note below

Why community-1 is required. In pyannote-audio 4.x, building a SpeakerDiarization pipeline loads a PLDA model unconditionally, then only uses it when clustering is VBxClustering. This pipeline uses AgglomerativeClustering, so the PLDA is fetched and ignored — but the download still happens, so you need access. Expect a first-run download you won't actually use.

Usage

from pyannote.audio import Pipeline
import torch

pipeline = Pipeline.from_pretrained(
    "daanshil/pyannote-seg-afr-childspeech-ft", token="hf_..."
)
pipeline.to(torch.device("cuda"))  # optional

output = pipeline("recording.wav")

# pyannote 4.x returns a DiarizeOutput wrapper — unwrap to get the Annotation
diarization = getattr(output, "speaker_diarization", output)

for turn, _, speaker in diarization.itertracks(yield_label=True):
    print(f"{turn.start:.2f}s - {turn.end:.2f}s: {speaker}")