Try using something like Pyodide or Coldbrew to get wav2vec running in the browser.

I was playing around with your API -- which is a fantastic little tool -- and I came across an issue where converting a .wav to .svg using a batch script fails to create a usable .svg. I've attached the batch script I'm using (in .txt format). As a side note, this does create an .svg file, it just does not have usable data. Any insight into this issue would be great. Thanks in advance.

wavTosvg.txt

import fairseq
import torch
import torchaudio
import matplotlib.pyplot as plt
from scipy.spatial.distance import cdist

Downloaded from https://dl.fbaipublicfiles.com/fairseq/wav2vec/xlsr_53_56k.pt on 2021-01-27

wav2vec2_checkpoint_path = "xlsr_53_56k.pt"

Code from facebookresearch/fairseq#3134 (comment)

checkpoint = torch.load(wav2vec2_checkpoint_path)
wav2vec2_encoder = fairseq.models.wav2vec.Wav2Vec2Model.build_model(checkpoint['cfg']['model'])
wav2vec2_encoder.load_state_dict(checkpoint['model'])

q_dat, q_sr = torchaudio.load("hello.wav")
r_dat, r_sr = torchaudio.load("goodbye-hello-goodbye.wav")

Resample to 16 kHz

q_dat = torchaudio.transforms.Resample(q_sr, 16000)(q_dat)
r_dat = torchaudio.transforms.Resample(r_sr, 16000)(r_dat)

Extract features

query_wav2vec2 = wav2vec2_encoder(q_dat, features_only=True, mask=False)['x'].detach().numpy().squeeze()
reference_wav2vec2 = wav2vec2_encoder(r_dat, features_only=True, mask=False)['x'].detach().numpy().squeeze()

Calculate distance matrix

qr_dists_w2v2 = cdist(query_wav2vec2, reference_wav2vec2, 'euclidean', V = None) # Calculate distance matrix
qr_dists_w2v2 = ((qr_dists_w2v2 - qr_dists_w2v2.min())/(qr_dists_w2v2.max() - qr_dists_w2v2.min())) # Normalized to [0, 1]

Plot distance matrix

plt.imshow(qr_dists_w2v2, interpolation='none')
plt.show()

Hi, wav2vec team.

I received "TypeError: "NoneType" object is not subscriptable" as in the screenshot attached when I converted some wav files while others seem to be OK.

Can you pls give some ideas what caused it?

Thank you.