Speech Recognition

Intro

The fastai library simplifies training fast and accurate neural nets using modern best practices. See the fastai website to get started. The library is based on research into deep learning best practices undertaken at fast.ai, and includes “out of the box” support for vision, text, tabular, and collab (collaborative filtering) models.

Dataset

Grab data from TensorFlow Speech Commands (2.3 GB):

commands_path = "SPEECHCOMMANDS"
audio_files = get_audio_files(commands_path)
length(audio_files$items)
# [1] 105835

Preprocess

Prepare dataset and put into data loader:

DBMelSpec = SpectrogramTransformer(mel=TRUE, to_db=TRUE)
a2s = DBMelSpec()
crop_4000ms = ResizeSignal(4000)
tfms = list(crop_4000ms, a2s)
auds = DataBlock(blocks = list(AudioBlock(), CategoryBlock()),  
                 get_items = get_audio_files, 
                 splitter = RandomSplitter(),
                 item_tfms = tfms,
                 get_y = parent_label)

audio_dbunch = auds %>% dataloaders(commands_path, item_tfms = tfms, bs = 20)

See batch:

audio_dbunch %>% show_batch(figsize = c(15, 8.5), nrows = 3, ncols = 3, max_n = 9, dpi = 180)

Model

Before fitting, 3 channels to 1 channel:

torch = torch()
nn = nn()

alter_learner = function(learn, channels = 1L) {
  learn$model[0][0][['in_channels']] %f% channels
  learn$model[0][0][0][['weight']] %f% torch$nn$parameter$Parameter(
  (learn$model[0][0][0]$weight %>% narrow('[:,1,:,:]'))$unsqueeze(1L)
)
}

learn = Learner(audio_dbunch, xresnet18(pretrained = FALSE), nn$CrossEntropyLoss(), metrics=accuracy)

nnchannels = audio_dbunch %>% one_batch() %>% .[[1]] %>% .$shape %>% .[1]

alter_learner(learn, nnchannels)

Add callbacks

Weights and biases could be save and visualized on wandb.ai:

# login for the 1st time then remove it
login("API_key_from_wandb_dot_ai")
init(project='R')
wandb: Currently logged in as: henry090 (use `wandb login --relogin` to force relogin)
wandb: Tracking run with wandb version 0.10.8
wandb: Syncing run macabre-zombie-2
wandb: ⭐️ View project at https://wandb.ai/henry090/speech_recognition_from_R
wandb: 🚀 View run at https://wandb.ai/henry090/speech_recognition_from_R/runs/2sjw3juv
wandb: Run data is saved locally in wandb/run-20201030_224503-2sjw3juv
wandb: Run `wandb off` to turn off syncing.

Conclusion

Now we can train our model:

learn %>% fit_one_cycle(3, lr_max=slice(1e-2), cbs = list(WandbCallback()))
epoch   train_loss   valid_loss   accuracy   time 
------  -----------  -----------  ---------  -----
epoch   train_loss   valid_loss   accuracy   time 
------  -----------  -----------  ---------  -----
WandbCallback requires use of "SaveModelCallback" to log best model
0       0.590236     0.728817     0.787121   04:18 
WandbCallback was not able to get prediction samples -> wandb.log must be passed a dictionary
1       0.288492     0.310335     0.908490   04:19 
2       0.182899     0.196792     0.941088   04:10 

See beautiful dashboard here:

https://wandb.ai/henry090/speech_recognition_from_R/runs/2sjw3juv?workspace=user-henry090