Preprint
Leveraging Self-Supervised Models for Automatic Whispered Speech Recognition
Aref Farhadipour et al., arXiv (Electrical Engineering and Systems Science, Audio and Speech Processing), 2024
Read the originalRead on October 16, 2026
- What it is
- A 2024 preprint that tested OpenAI's Whisper model on whispered speech, then fine-tuned a WavLM model on whispered and normal speech from two English datasets (Singaporean and Irish dialects).
- How much weight it carries
- Peer-review status not confirmed for this version. The test data are two research corpora of read speech in English dialects, not everyday dictation, and the Whisper figure reflects the model version used in 2024. Treat it as evidence that whispered speech is hard for a general model, not as a measure of current dictation software.
Facts taken from this source
- In a 2024 baseline test, the OpenAI Whisper model had a word error rate of 18.8 percent on whispered speech.
“achieving a Word Error Rate (WER) of 18.8% and a Character Error Rate (CER) of 4.24% on whispered speech” (Abstract)