Performance issue on the embeddings step

[kenenbek]

Tested versions

pyannote.audio-4.0.1
pyannote.audio-4.0.2

System information

Description: Ubuntu 24.04.3 LTS Release: 24.04

Issue description

I used a pyannote on a five mins long audio. Segmentation, speaker_counting take just 2 seconds. But embeddings takes a really long time. My GPU is RTX-4070.

My versions:

pyannote.audio-4.0.2
torchcodec-0.7.0
torch '2.8.0+cu128'
numpy-2.3.5

segmentation     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100% 0:00:02
speaker_counting ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100% 0:00:00
embeddings       ━━━━━━━━━━━━━━━━━━━━╺━━━━━━━━━━━━━━━━━━━  50% 0:08:35

Minimal reproduction example (MRE)

https://gist.github.com/kenenbek/22c66fb788487f64ce8ecacfe1626535

[clement-pages]

Hey @kenenbek, you can try to speed up the processing by preloading the audio, then call the pipeline:

pipeline({"waveform": (channel, time) torch.Tensor, "sample_rate": audio_sample_rate})

[kenenbek]

@clement-pages It is unbelievable, before it was:

✓ TEST COMPLETE!
⏱ Time taken: 17.57 minutes (1053.9 seconds)

Now it is:

✓ TEST COMPLETE!
⏱ Time taken: 0.83 minutes (49.8 seconds)

You saved my week!

[Karen86Tonoyan]

Woow congratulation and come tu my web Karen86Tonoyan czw., 27 lis 2025, 04:57 użytkownik Kenenbek R ***@***.***> napisał:

…

*kenenbek* left a comment (pyannote/pyannote-audio#1955) <#1955 (comment)> @clement-pages <https://github.com/clement-pages> It is unbelievable, before it was: ✓ TEST COMPLETE! ⏱ Time taken: 17.57 minutes (1053.9 seconds) Now it is: ✓ TEST COMPLETE! ⏱ Time taken: 0.83 minutes (49.8 seconds) You saved my week! — Reply to this email directly, view it on GitHub <#1955 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BUHB4E6XDZEK7LLHG6BJZIL36ZZBRAVCNFSM6AAAAACMZBLVGOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTKOBUGEYDINJQGQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[crysingzz]

@clement-pages It is unbelievable, before it was:

✓ TEST COMPLETE!
⏱ Time taken: 17.57 minutes (1053.9 seconds)

Now it is:

✓ TEST COMPLETE!
⏱ Time taken: 0.83 minutes (49.8 seconds)

You saved my week!

hello! can u share ur new full code please? :D

[kenenbek]

@crysingzz here is full code

def diarize_audio(audio_file, num_speakers=None, device=None):
    
    # Set device
    if device is None:
        device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    print(f"Using device: {device}")

    # Load pipeline
    print("Loading diarization pipeline...")
    pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-community-1")
    #pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1")
    pipeline.to(device)


    # --- PRELOADING AUDIO ---
    print(f"Preloading audio from {audio_file}...")
    # torchaudio.load returns: waveform (channel, time), sample_rate (int)
    waveform, sample_rate = torchaudio.load(audio_file)

    # Construct the dictionary expected by pyannote
    # Note: Keep waveform on CPU; the pipeline moves specific chunks to GPU internally.
    audio_in_memory = {
        "waveform": waveform,
        "sample_rate": sample_rate
    }

    print(f"Performing diarization on {audio_file}...")
    with ProgressHook() as hook:
        if num_speakers:
            diarization = pipeline(audio_in_memory, hook=hook, num_speakers=num_speakers)
        else:
            diarization = pipeline(audio_in_memory, hook=hook)

    print("Diarization complete!")
    return diarization

[crysingzz]

def diarize_audio(audio_file, num_speakers=None, device=None):

thanks!

[crysingzz]

@clement-pages It is unbelievable, before it was:

✓ TEST COMPLETE!
⏱ Time taken: 17.57 minutes (1053.9 seconds)

Now it is:

✓ TEST COMPLETE!
⏱ Time taken: 0.83 minutes (49.8 seconds)

You saved my week!

what is duration of ur audio?
guys, have u methods how speed up pyannote diarizarion? i need diarize my 50min audio faster than 100 seconds...