Whisper Khmer v9 (aug02)
Fine-tuned openai/whisper-large-v3-turbo for Khmer (Cambodian) speech recognition, using a
custom merged BPE tokenizer (whisper-khmer-tokenizer-v3, content-first vocab layout).
Trained with online audio augmentation.
- Base model:
openai/whisper-large-v3-turbo - Tokenizer: custom merged Khmer BPE (57,908 tokens)
- Training: online audio augmentation (11 techniques),
augment_prob=0.3, noisy eval - Clean eval WER: ~22% at step 1200
- Vocab layout: content-first — custom BPE tokens precede special tokens, so the
standard
transformersASR pipeline decodes them correctly.
Usage
from transformers import pipeline
pipe = pipeline("automatic-speech-recognition",
model="1morecupofhottea/whisper-turbo-khmer-v9",
chunk_length_s=30, device=0)
result = pipe("path/to/audio.wav")
print(result["text"]) # khmercut-segmented output with ';' word-group separators
The model outputs khmercut-segmented text (words separated by spaces, word groups by ;),
which matches the training data format.