File size: 1,571 Bytes
7032ae4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
"""
Standalone example: turn a raw audio file into a HuBERT discrete-unit
sequence, using only the two model folders in this repo.

Usage:
    python extract_units.py path/to/audio.wav

Requirements:
    pip install transformers torchaudio joblib numpy
"""
import sys
import numpy as np
import torch
import torchaudio
import joblib
from transformers import AutoModel

TARGET_SR = 16000
LAYER = 6  # must match the layer this repo's k-means model was fit on -- see unit-discovery-model/config.txt


def load_hubert(checkpoint_dir):
    model = AutoModel.from_pretrained(checkpoint_dir)
    model.eval()
    return model


def extract_features(model, wav_path, layer):
    waveform, sr = torchaudio.load(wav_path)
    if waveform.shape[0] > 1:
        waveform = waveform.mean(dim=0, keepdim=True)
    if sr != TARGET_SR:
        waveform = torchaudio.functional.resample(waveform, sr, TARGET_SR)
    with torch.no_grad():
        out = model(waveform, output_hidden_states=True)
        feats = out.hidden_states[layer].squeeze(0).numpy()
    return feats


def dedup(sequence):
    out = []
    for x in sequence:
        if not out or out[-1] != x:
            out.append(x)
    return out


if __name__ == "__main__":
    audio_path = sys.argv[1]
    hubert_model = load_hubert("../hubert-model")
    km = joblib.load("../unit-discovery-model/kmeans.joblib")

    feats = extract_features(hubert_model, audio_path, LAYER)
    unit_ids = km.predict(feats).tolist()
    deduped = dedup(unit_ids)

    print("Raw units:   ", unit_ids)
    print("Deduped units:", deduped)