pyannote_speaker_diarization_pipeline_with_gpu_inference.py

1import torch
2from pyannote.audio import Pipeline
3
4# 1. Initialize the pipeline from Hugging Face
5# Note: You must accept the user conditions on Hugging Face for:
6# - pyannote/speaker-diarization-3.1
7# - pyannote/segmentation-3.0
8pipeline = Pipeline.from_pretrained(
9    "pyannote/speaker-diarization-3.1",
10    use_auth_token="HUGGINGFACE_ACCESS_TOKEN_HERE"
11)
12
13# 2. Send the pipeline to GPU if available
14device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
15pipeline.to(device)
16
17# 3. Apply the pipeline to an audio file
18diarization = pipeline("audio.wav")
19
20# 4. Print the result (start time, end time, and speaker label)
21for turn, _, speaker in diarization.itertracks(yield_label=True):
22    print(f"start={turn.start:.1f}s stop={turn.end:.1f}s speaker_{speaker}")