Speaker Diarizations get_segmentations() raises for several input type variants

[bschreck]

Tested versions

• tested in 3.1

System information

macOs 13.6 - pyannote 3.1 - M2 air

Issue description

2 variants that according to the docstring should be correct inputs instead raise a ValueError:

pipeline=Pipeline.from_pretrained(
            "pyannote/speaker-diarization-3.1", use_auth_token=os.environ["HF_API_KEY"]
        )

audio = Audio()({'waveform': waveform, 'sample_rate': sample_rate}) # waveform is 2d numpy array
segmentations = pipeline.get_segmentations(audio)  # raises
segmentations = pipeline.get_segmentations({'waveform': waveform, 'sample_rate': sample_rate})  # raises
segmentations = pipeline.get_segmentations({'waveform': torch.from_numpy(waveform), 'sample_rate': sample_rate})  # succeeds

Minimal reproduction example (MRE)

see above