SLaM_DeepSpeech

An implementation of DeepSpeech's pre-trained English model for use in the SLaM lab at the University of Florida.

Package Versions

• deepspeech 0.6.1
• jiwer 1.3.2
• numpy 1.17.3
• PyAudio 0.2.11
• scipy 1.3.1
• tensorflow 1.14.0

Using the SLaM DeepSpeech Interface

This interface is intended to make using DeepSpeech pre-trained models more compatible with large batches of audio files. It also includes build-in error calculations. Currently, Word Error Rate is implemented using the jiwer Package.

Command Line Functionality

• -h help
• -i inputdir, the path to a directory containing audio files; audio files should be in .wav format; transcriber will skip any files without a .wav extension
• -o outputfile, the path to an empty .txt file where the results are written
• -m model directory, the path to the directory containing the DeepSpeech pre-trained model; should contain
• -r error type, current options: word phone ; if -r is not included, error will not be calculated
• -g ground_truth, the path to a .txt file containing the intended transcriptions of the audio files; used to calculate error rate; transcriptions should be in sorted order, separated by new lines
• -a lm_alpha, the relative weight of language model vs. Correctionist Temporal Classification (CTC)
• -b lm_beta, considers more words

Example

python3 SLaM_DeepSpeech -i $HOME/deepspeech-venv/SLaM_DeepSpeech/SLaM_DeepSpeech/temp -o $HOME/deepspeech-venv/SLaM_DeepSpeech/SLaM_DeepSpeech/temp/output.txt -m $HOME/deepspeech-venv/deepspeech-0.6.1-models -r word -g $HOME/deepspeech-venv/SLaM_DeepSpeech/SLaM_DeepSpeech/temp/gt.txt