whisper-turbo-khmer-v9, CTranslate2 int8
A CTranslate2 int8 conversion of
1morecupofhottea/whisper-turbo-khmer-v9
(MIT) at revision f8644583afbf77707a0fc75eef36f80d116eeb1f, for Khmer speech recognition on a CPU.
The model and its training are the original authors' work; this repository only changes its format.
- Measured: CER 0.1645 on 150 FLEURS
km_khutterances (the original in float16: 0.148), p50 5.62 s per utterance on 4 CPU threads, by the Vatana AI Platform's own evaluation (serving/scripts/eval-khmer-ct2.py). - Decoding: this fine-tune keeps its Khmer tokens above
<|endoftext|>, and faster-whisper's text assembly drops them (CER 0.97). Decode with CTranslate2 directly and remove control tokens by name, asserving/spaces/khmer-asr/ct2_engine.pyin the platform does. - Made with:
ct2-transformers-converter --quantization int8(serving/scripts/convert-khmer-ct2.sh).