hi-vaani (CTranslate2)
A CTranslate2 conversion of ARTPARK-IISc/whisper-large-v3-vaani-hindi for faster-whisper, as downloaded by LiveWhisper, a voice dictation app for Windows.
All credit for the model goes to its original authors. This repository only
changes the file format (float16 weights, plus tokenizer.json and
preprocessor_config.json that faster-whisper needs). The licence is the
original's: apache-2.0.
For Devanagari output. Large-v3 sized, 3 GB converted.
Measured in LiveWhisper
| metric | result |
|---|---|
| FLEURS test WER, native script | 10.3% (large-v3 26.5%) |
| FLEURS test CER, native script | 3.0% (large-v3 8.7%) |
| code-switched Hindi spelled exactly | 96.0% (large-v3 94.0%) |
from faster_whisper import WhisperModel
model = WhisperModel("itsskofficial/livewhisper-hi-vaani", device="cuda", compute_type="int8_float16")