Skip to content

Latest commit

 

History

History
33 lines (17 loc) · 951 Bytes

README.md

File metadata and controls

33 lines (17 loc) · 951 Bytes
Raw

NLP-speech-to-text

Convert speech to text using HuggingFace, comparing Wav2Vec2 versus OpenAI Whisper

Data

Speech samples included a subset of sentences recorded for this study:

Reuter, T., Sullivan, M., & Lew-Williams, C. (2021). Look at that: Spatial deixis reveals experience- related changes in prediction. Language Acquisition. https://doi.org/10.1080/10489223.2021.1932905

Audio for lab-based experiments are very clean. So this should be an easy transcription task.

Conclusion

IMO, Whisper beats Wav2Vec2 in at least 3 ways:

  1. More performant.

  • Transcribed 20% faster.

  • Future enhancements could increase speed.

  1. More accurate.

  • Transcribed "apple" versus "apples" correctly.

  • Spelled "doggies" correctly as "doggies", not as "DOGGIYS".

  1. More nuanced.

  • Transcribed 3 sentences with emphatic punctuation (! instead of .)

  • Punctuation indicates emphasis and emotion, useful for downstream sentiment analysis.