Files
ea18bc70d8 Align audio CI and releases; require Python 3.11 (#2081)
* feat(cli): add --subfolder option to benchmark command (#2063)
* fix(utils): fix `SpeakerDiarizationMixin.optimal_mapping` when `reference` is a mapping
* fix(task): fix `balance` option crashing training with `KeyError: 'metadata'` (#2067)
* fix(io): fix `Audio.crop` raising when an in-memory waveform is cropped up to its very end (#2068)
* fix(model): make Model.default_activation return identity for unsupported tasks
* fix: compute sample duration from preloaded waveform
* fix(pipeline): omit unset runtime options from model references (#2080)

Co-authored-by: abeeranajam31 <128603901+abeeranajam31@users.noreply.github.com>
Co-authored-by: antoinelaurent <antoinelaurent@me.com>
Co-authored-by: Kayvan Zahiri <123409429+Kayvan-Zahiri@users.noreply.github.com>
Co-authored-by: Mohammad Hijjawi <51020382+MohammadHijjawi97@users.noreply.github.com>
2026-10-22 15:07:01 +02:00

19 lines
542 B
Python

# Copyright (c) 2025- pyannoteAI
import torch
from pyannote.audio.models.embedding.wespeaker import WeSpeakerResNet34
def test_fbank_only():
model = WeSpeakerResNet34().eval()
waveforms = torch.randn(1, 1, 16000)
with torch.inference_mode():
expected = model.compute_fbank(waveforms)
model.fbank_only = True
features = model(waveforms)
torch.testing.assert_close(features, expected)
model.fbank_only = False
embeddings = model(waveforms)
assert embeddings.shape == (1, 256)