Automatic Speech Recognition
Welsh
whispercpp
File size: 1,720 Bytes
36a7fe5
 
 
 
 
 
 
 
 
 
7d2404e
 
36a7fe5
 
 
 
 
6820fe3
 
 
 
 
 
 
 
 
 
 
 
36a7fe5
 
6820fe3
 
36a7fe5
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
---
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](https://huggingface.co/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](https://github.com/ggerganov/whisper.cpp/tree/master/models#fine-tuned-models). 

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](https://huggingface.co/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](https://github.com/ggerganov/whisper.cpp?tab=readme-ov-file#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`