A simple waveform segmentator using OpenAI's Whisper
Recently proposed deep-learning based speech processing models require batch processing due to limited memory space. To train the models, generally, we segment a long waveform into utterance-level segments. In the wild, however, most of the speech dataset collected from record studio or web-scrapping contains long waveform which consists of multiple utterances, and makes inappropriate to train them. Manual segmentation and annotation are needed in some cases, but time-inefficient.
We propose SpeechDatasetSplitter to automate the segmentation and annotation process. The proposed SpeechDatasetSplitter is built on top of the OpenAI's Whisper, and capable to sophisticatedly segment long-waveform into multiple utterances. Moreover, automated annotation and validation are also possible for efficient and effective data preprocessing.
- strongly recommend to run under Anaconda environment
- install pre-requisite via following command.
pip install -r requirements.txt
- upload waveform in the
samplesdirectory. - run script(
run.py) via following command
python run.py
- check the segmented waveforms in
resultsdirectory - you may check the error report named
result.csv.
If any question, please email to mskang1478@gmail.com.