End-to-End Sentence-Classification-on-FSCs-Schmalz

End-to-End Sentence Classification based on the Fluent Speech Commands Dataset

Using the opensource Fluent Speech Commands dataset (available at https://fluent.ai/fluent-speech-commands-a-dataset-for-spoken-language-understanding-research), we consider an End-2-End phrase classification task. The dataset contains 30,043 utterances and 248 possible sentences. The utterances are pronounced by both native and non-native English speakers and include phrases like "turn off the lights in the kitchen" or "heat up in the living room", while the possible sentences are intended to define the action, object and location of them, for example "deactivate, lights, kitchen" or "increase, heat, living room" for the previously mentioned phrases.

In order to address the task we adopt a neural framework that classifies the utterances to the possible sentences. The model used in the experiment is a Time Convoluted Network (TCN, available at https://github.com/asteroid-team/asteroid from https://github.com/popcornell/OSDC. The network receives as input fixed length sequences of 40 Mel filter-banks. The signal length is limited to 4 seconds or 64000 samples. Filter banks are computed on 20ms window with 10ms hop size, resulting in 400 frames.

Pre-processing

• cutfiles.py:

fcut,index= librosa.effects.trim(f,frame_length=2098, hop_length=562)

./fluent_speech_commands_dataset/ identifies the folder in which the data are to be found.

Training

• main.py :

python3.6 main.py -n TCN -m models/tcn_b5r2.pkl -b 5 -r 2 -lr 0.001 -e 100

• -n: type of net
• -m: model
• -b: number of blocks of the TCN network
• -r: number of repeats of the TCN network
• -lr: learning rate
• -e: number of epochs

Evaluation

• evaluation.py:

python3.6 evaluation.py -n TCN -m models/tcn_b5r2.pkl -b 5 -r 2 

Using the training parameters suggested above, the obtained results should be 0.816870 accuracy on the validation set and 0.934880 accuracy on the evaluation set.