Request for pre-TFLite ACFT-KO large-v3-turbo checkpoint for CTranslate2 testing
Hi, thank you for sharing the ACFT-KO models.
I am testing faster-whisper with large-v3-turbo on CPU using short Korean telephone recordings, typically around 3–4 seconds long.
I experimented with reducing the encoder input length instead of padding every utterance to 30 seconds. This significantly reduced the processing time, but Korean recognition accuracy also dropped substantially. I would therefore like to test whether ACFT-KO can maintain better accuracy with shorter encoder inputs.
Would you be willing to share the fine-tuned ACFT-KO large-v3-turbo checkpoint from before the TFLite conversion? Hugging Face Transformers format would be ideal, including the model weights (such as safetensors), config.json, and any required tokenizer or preprocessing configuration.
My goal is to convert the checkpoint to CTranslate2 and compare transcription accuracy and latency on the same recordings using 5-second and 10-second encoder windows.
I am specifically looking for the fine-tuned training checkpoint rather than an FP32 TFLite export. Any information about custom model code, positional embeddings, or inference changes required for short audio contexts would also be very helpful.
If the checkpoint cannot currently be shared, are there any plans to release it in the future?
Thank you for your time and for your work on this project.
I apologize, but I am unable to provide the files.
They were deleted due to storage limitations in the workspace.
Since the training process is not difficult, I would recommend trying it out yourself.