|
|
--- |
|
|
library_name: transformers |
|
|
language: |
|
|
- twi |
|
|
license: apache-2.0 |
|
|
base_model: openai/whisper-tiny |
|
|
tags: |
|
|
- custom-dataset |
|
|
- local-dataset |
|
|
- whisper |
|
|
- generated_from_trainer |
|
|
metrics: |
|
|
- wer |
|
|
model-index: |
|
|
- name: T6-Whisper-FineTuned-DL-Twi |
|
|
results: [] |
|
|
--- |
|
|
|
|
|
<!-- This model card has been generated automatically according to the information the Trainer had access to. You |
|
|
should probably proofread and complete it, then remove this comment. --> |
|
|
|
|
|
# T6-Whisper-FineTuned-DL-Twi |
|
|
|
|
|
This model is a fine-tuned version of [openai/whisper-tiny](https://huggingface.co/openai/whisper-tiny) on the Twi-native Ghanaian language. dataset. |
|
|
It achieves the following results on the evaluation set: |
|
|
- Loss: 0.0063 |
|
|
- Wer: 23.4562 |
|
|
- Cer: 21.7611 |
|
|
|
|
|
## Model description |
|
|
|
|
|
T6-Whisper-FineTuned-DL-Twi is a fine-tuned version of openai/whisper-tiny focused specifically on the Twi language, a widely spoken native language in Ghana. This model adapts Whisper’s multilingual speech recognition capabilities to better understand and transcribe Twi speech, especially in financial contexts. |
|
|
|
|
|
It was developed as part of a project to support accessibility in financial systems, aiming to make digital financial services more inclusive for Ghanaian communities that primarily communicate in Twi. |
|
|
|
|
|
## Intended uses & limitations |
|
|
|
|
|
Intended uses: |
|
|
- Automatic Speech Recognition (ASR) for Twi and English-Twi mixed audio. |
|
|
- Enhancing voice interfaces in fintech platforms (e.g., mobile banking, customer support). |
|
|
- Increasing accessibility for low-literate or visually impaired users in financial contexts. |
|
|
- Supporting research in code-switched speech and low-resource African languages. |
|
|
|
|
|
Limitations: |
|
|
- May not perform optimally outside the financial domain (e.g., health or legal speech). |
|
|
- Performance can degrade in noisy environments or with heavy accents not represented in the training data. |
|
|
- While it handles code-switching, rapid or highly irregular switches may still reduce accuracy. |
|
|
- Based on the Whisper-tiny model, which is optimized for speed and size, not peak performance. |
|
|
|
|
|
## Training and evaluation data |
|
|
|
|
|
The model was fine-tuned using a custom dataset containing Twi and English-Twi code-switched audio, primarily from the financial domain. This includes content like: |
|
|
|
|
|
- Mobile money instructions |
|
|
- Banking app voice interactions |
|
|
- Financial literacy radio shows and interviews |
|
|
- Call center conversations involving customer queries |
|
|
- Dataset size: ~ 50 hours |
|
|
- Language mix: Twi + English (code-switched) |
|
|
- Transcript quality: Manually verified by native speakers |
|
|
- Train/validation split: [e.g., 80/20] |
|
|
|
|
|
|
|
|
## Training procedure |
|
|
|
|
|
### Training hyperparameters |
|
|
|
|
|
The following hyperparameters were used during training: |
|
|
- learning_rate: 1e-05 |
|
|
- train_batch_size: 16 |
|
|
- eval_batch_size: 8 |
|
|
- seed: 42 |
|
|
- optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments |
|
|
- lr_scheduler_type: linear |
|
|
- lr_scheduler_warmup_steps: 500 |
|
|
- training_steps: 4000 |
|
|
- mixed_precision_training: Native AMP |
|
|
|
|
|
### Training results |
|
|
|
|
|
| Training Loss | Epoch | Step | Validation Loss | Wer | Cer | |
|
|
|:-------------:|:------:|:----:|:---------------:|:-------:|:-------:| |
|
|
| 0.025 | 0.6333 | 1000 | 0.0285 | 27.9879 | 21.3775 | |
|
|
| 0.0083 | 1.2666 | 2000 | 0.0094 | 20.4318 | 17.7329 | |
|
|
| 0.0058 | 1.8999 | 3000 | 0.0072 | 19.5177 | 17.5028 | |
|
|
| 0.0012 | 2.5332 | 4000 | 0.0063 | 23.4562 | 21.7611 | |
|
|
|
|
|
|
|
|
### Framework versions |
|
|
|
|
|
- Transformers 4.48.0.dev0 |
|
|
- Pytorch 2.5.1+cu121 |
|
|
- Datasets 3.2.0 |
|
|
- Tokenizers 0.21.0 |
|
|
|
