---
license: other
language:
- gd
base_model:
- openai/whisper-large-v3-turbo
tags:
- audio
- automatic-speech-recognition
library_name: ctranslate2
---

# Fine-Tuned Whisper Large-v3 Turbo for Scottish Gaelic (CTranslate2)

This repository contains the conversion of a fine-tuned **Whisper Large-v3 Turbo** model for **Scottish Gaelic** to the [CTranslate2](https://github.com/OpenNMT/CTranslate2) model format.

This model is based on the research presented in the paper [*"Practitioner’s Guide to Building SR Models for Low-Resource Languages: Case Study on Scottish Gaelic"* (Klejch, Lamb, Bell, 2025)](https://www.isca-archive.org/interspeech_2025/klejch25_interspeech.html).

This model can be used in CTranslate2 or projects based on CTranslate2 such as [faster-whisper](https://github.com/systran/faster-whisper).

## Benchmarks & Results

The following results are based on the experiments conducted in the associated paper using the Whisper Turbo architecture. The Word Error Rate (WER) is reported across four test sets: **CLTW**, **News**, **BBC**, and **PC** (Prògram Choinnich).

| Model Configuration | CLTW | News | BBC | PC | **Avg WER** |
|:--- | :---: | :---: | :---: | :---: | :---: |
| Whisper Turbo (LoRA 64) | 25.7 | 23.1 | 31.5 | 27.4 | **26.9** |
| Whisper Turbo (LoRA 512) | 20.8 | 19.2 | 26.9 | 21.2 | **22.0** |
| Whisper Turbo (LoRA 512 + semi-supervised data) | 19.4 | 15.0 | 23.5 | 18.1 | **19.0** |
| **Whisper Turbo (Best Config)**<br>*(full fine-tuning + semi-supervised data)* | **12.8** | **10.0** | **15.9** | **13.5** | **13.0** |

*Note: The results above reflect the performance detailed in Table 3 of the paper. Performance may vary depending on the specific checkpoint version and quantization used.*

## Example Usage

```python
from faster_whisper import WhisperModel

# Load the model
model = WhisperModel("eist-edinburgh/whisper-large-v3-turbo-gaelic-ct2-v2")

# Transcribe a Scottish Gaelic audio file
segments, info = model.transcribe("gaelic_audio.mp3", language="cy")

for segment in segments:
    print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))
```

## Conversion Details

The original model was converted with the following command:

```bash
ct2-transformers-converter --model deepdml/whisper-large-v3-turbo --output_dir faster-whisper-large-v3-turbo \
    --copy_files tokenizer.json preprocessor_config.json --quantization float16
```

Note that the model weights are saved in FP16. This type can be changed when the model is loaded using the [`compute_type` option in CTranslate2](https://opennmt.net/CTranslate2/quantization.html).

## Citation

If you use this model or the associated results, please cite the original paper:

```bibtex
@inproceedings{klejch25_interspeech,
  title     = {{A Practitioner’s Guide to Building ASR Models for Low-Resource Languages: A Case Study on Scottish Gaelic}},
  author    = {Ondřej Klejch and William Lamb and Peter Bell},
  year      = {2025},
  booktitle = {{Interspeech 2025}},
}
```

## More Information

**For more information about the base architecture, see the [Whisper Large v3 Turbo model card](https://huggingface.co/openai/whisper-large-v3-turbo).**

## License

The Gaelic Automatic Speech Recognition model and API were developed by the University of Edinburgh with funding from the Scottish Government. They are licensed under the Open Government License v3.0. If you would like to engage with the research team behind the Gaelic Automatic Speech Recognition project, please contact Professor Will Lamb, Personal Chair in Gaelic Ethnology and Linguistics at the University of Edinburgh (w.lamb@ed.ac.uk).