license: apache-2.0
datasets:
- techiaith/banc-trawsgrifiadau-bangor
- techiaith/commonvoice_18_0_cy
language:
- cy
base_model:
- openai/whisper-base
pipeline_tag: automatic-speech-recognition
tags:
- whispercpp
whisper-base-ft-btb-cv-cy-cpp
This model is a version of the openai/whisper-base model, fine-tuned with transcriptions of Welsh language spontaneous speech from Banc Trawsgrifiadau Bangor (btb) dataset, as well as read speech from Welsh Common Voice version 18 (cv) for additional training, and then converted for use in whisper.cpp.
Whispercpp is a C/C++ port of Whisper that provides high performance inference on hardware such as desktops, laptops and mobile devices, thus giving an offline option.
The model is a smaller in size to the corresponding model for hosting on cloud GPU based infrastructure techiaith/whisper-large-v3-ft-btb-cv-cy and thus not as accurate.
It achieves the following WER results for transcribing Welsh language spontaneous speech:
- WER: 62.76
- CER: 27.70
Usage
whispercpp makes it easy to use models in many platforms and applications. See the 'examples' folder in the whispercpp github repo for more information and example code.
To get quickly started with whispercpp's basic usage however, follow the 'Quick Start' but download this model with the following command:
$ wget https://huggingface.co/techiaith/whisper-base-ft-btb-cv-cy-cpp/resolve/main/ggml-model.bin