---
library_name: mlx
license: apache-2.0
language:
- km
pipeline_tag: automatic-speech-recognition
datasets:
- seanghay/km-speech-corpus
- seanghay/khmer_mwpt_speech
tags:
- Khmer
- mlx
base_model: openai-whisper-tiny
model-index:
- name: whisper-tiny-khmer-mlx-fp32 by Kimang KHUN
  results:
  - task:
      type: automatic-speech-recognition
      name: Speech Recognition
    dataset:
      name: test split of "km_kh" in google/fleurs
      type: google/fleurs
    metrics:
    - type: wer
      value: 73.5%
      name: test
  - task:
      type: automatic-speech-recognition
      name: Speech Recognition
    dataset:
      name: train split of "SLR42" in openslr/openslr
      type: openslr/openslr
    metrics:
    - type: wer
      value: 56.4%
      name: test
---

# whisper-tiny-khmer-mlx-fp32
This model was converted to MLX format from [`openai-whisper-tiny`](https://github.com/openai/whisper), then fine-tined to Khmer language using two datasets:
- [seanghay/khmer_mpwt_speech](https://huggingface.com/datasets/seanghay/khmer_mpwt_speech)
- [seanghay/km-speech-corpus](https://huggingface.com/datasets/seanghay/km-speech-corpus)

It achieves the following __word error rate__ (`wer`) on 2 popular datasets:
- 73.5% on `test` split of [google/fleurs](https://huggingface.co/datasets/google/fleurs) `km-kh`
- 56.4% on `train` split of [openslr/openslr](https://huggingface.co/datasets/openslr/openslr) `SLR42`

__NOTE__ MLX format is usable for M-chip series of Apple.

## Use with mlx
```bash
pip install mlx-whisper
```

Write a python script, `example.py`, as the following
```python
import mlx_whisper

result = mlx_whisper.transcribe(
    SPEECH_FILE_NAME,
    path_or_hf_repo="Kimang18/whisper-tiny-khmer-mlx-fp32",
    fp16=False
)
print(result['text'])
```
Then execute this script `example.py` to see the result.

You can also use command line in terminal
```bash
mlx_whisper --model Kimang18/whisper-tiny-khmer-mlx-fp32 --task transcribe SPEECH_FILE_NAME --fp16 False
```