ASR microservice developed for call center transcription service
Based on whisper-timestamped
The latest version of the Whisper model - v3, is used; service can operate on both GPU and CPU, but significantly slower on the latter. Prompt engineering was applied to improve the transcription results.
Transcription quality on Russian (source):
- WER - 0.2
- MER - 0.2
- WIL - 0.25