einstein8612/mlc-mlsw-melspects

MLCommons Multilingual Spoken Words Mel-Spectograms

This dataset contains all English words from the dataset available at MLCommons (or also available on huggingface). These audio files have been processed into Mel spectrograms for downstream usage in DCNNs or similar processes.

Dataset description

There's a total of 6624343 samples of Mel spectograms. There are a total of 38150 different words, the cls is the index of that word in alphabetical order. With every entry in the dataset there's the label, the Mel spectogram in a pickled PyTorch tensor in byte format, and a metadata JSON with the MD5 sum of the tensor representing the Mel spectogram and the actual word. The spectogram is in the shape of 128x32

You can view the Mel spectrogram of a word by viewing it through librosa.

Example

sr = 16000

import librosa
import torch
import io
import webdataset as wds
import matplotlib.pyplot as plt

def bytes_to_tensor(x):
    return torch.load(io.BytesIO(x))

shuffle_buffer = 1000
dataset = wds.WebDataset(urls).shuffle(shuffle_buffer).map_dict(pt=bytes_to_tensor, cls=int).to_tuple("pt", "cls")
spectogram, label = next(iter(dataset))

librosa.display.specshow(spectogram.numpy(), sr=sr, hop_length=512, y_axis='mel', fmax=8000, x_axis='time')
plt.title(f"Mel Spectrogram (y={label})")
plt.colorbar(format='%+2.0f dB')

Usage

As it's in the WebDataset format you can simply use it as is like above, or through a PyTorch dataloader (recommended).

Example (PyTorch dataloader)

import torch
import io
import webdataset as wds

def bytes_to_tensor(x):
    return torch.load(io.BytesIO(x))

shuffle_buffer = 1000
dataset = wds.WebDataset(urls).shuffle(shuffle_buffer).map_dict(pt=bytes_to_tensor, cls=int).to_tuple("pt", "cls")

loader = torch.utils.data.DataLoader(dataset, batch_size=256, num_workers=1)

for inputs, labels in loader:
    print(inputs.shape) # torch.Size([256, 128, 32])

Process behind dataset

The exact process was done using torchlibrosa using the following relevant code.

sr = 16000

class PowerToDB(nn.Module):
    def __init__(self, amin=1e-10, top_db=80.0):
        super(PowerToDB, self).__init__()

        self.amin = torch.tensor(amin)
        self.top_db = top_db

    def forward(self, S: torch.Tensor):
        refs = torch.amax(S, dim=(2,3))

        log_spec = 10.0 * torch.log10(self.amin.maximum(S)) - (10.0 * torch.log10(self.amin.maximum(refs))).unsqueeze(2).unsqueeze(3)
        log_spec = torch.maximum(log_spec, log_spec.max() - self.top_db)

        log_spec = log_spec[:, 0, :, :] # Remove channel dimension, always 1
        log_spec = log_spec.transpose(1, 2) # Swap mel bin and time axis

        return log_spec

feature_extractor = torch.nn.Sequential(
    tl.Spectrogram(
        pad_mode="constant",
        hop_length=512,
        win_length=None,
    ), tl.LogmelFilterBank(
        sr=sr,
        n_mels=128,
        is_log=False,
    ), PowerToDB(
        amin=1e-10,
        top_db=80.0,
    ))

This gives the same output as

sr = 16000

mel_spect = librosa.feature.melspectrogram(y=y, sr=sr, n_fft=2048, hop_length=512)
mel_spect = librosa.power_to_db(mel_spect, ref=np.max)

with some (negligible) degree of numerical error due to using PyTorch instead.