File size: 2,865 Bytes
f3e4b92
 
 
 
 
 
fc1a7ae
f3e4b92
 
f246440
 
 
 
 
 
 
e17606c
f3e4b92
90a7ef7
f246440
 
 
90a7ef7
f246440
 
 
 
 
 
 
90a7ef7
c2764d0
1038c44
f3e4b92
 
fc1a7ae
 
f3e4b92
fc1a7ae
f3e4b92
fc1a7ae
 
a925747
 
f3e4b92
4b77f7a
 
 
 
1397979
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f3e4b92
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fc1a7ae
f3e4b92
 
 
 
 
 
 
 
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
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
---
language:
- ml
license: apache-2.0
tags:
- whisper-event
- generated_from_trainer
datasets:
- mozilla-foundation/common_voice_11_0
- google/fleurs
- thennal/IMaSC
- thennal/ulca_ml
- thennal/msc
- thennal/indic_tts_ml
metrics:
- wer
base_model: openai/whisper-medium
model-index:
- name: Whisper Medium Malayalam - Thennal D K
  results:
  - task:
      type: automatic-speech-recognition
      name: Automatic Speech Recognition
    dataset:
      name: Common Voice 11.0
      type: mozilla-foundation/common_voice_11_0
      config: ml
      split: test
      args: ml
    metrics:
    - type: wer
      value: 11.49
      name: WER
---

<!-- 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. -->

# Whisper Medium Malayalam

This model is a fine-tuned version of [openai/whisper-medium](https://huggingface.co/openai/whisper-medium) on the Common Voice 11.0 dataset.
It achieves the following results on the evaluation set:
- WER: 38.6207
- CER: 7.3256

Note that Whisper's normalization has major issues for languages like Malayalam, so the above scores are evaluated without using normalization. 
With normalization (for a fair comparison with other models on this platform), the results are instead:
- WER: 11.49

[This Colab](https://colab.research.google.com/github/sanchit-gandhi/notebooks/blob/main/fine_tune_whisper.ipynb) can be used as a starting point to further finetune the model.
## Usage instructions
Given an audio sample `audio` (this can be anything from a numpy array to a filepath), the following code generates transcriptions:
```python
from transformers import pipeline, WhisperProcessor

processor = WhisperProcessor.from_pretrained("thennal/whisper-medium-ml")
forced_decoder_ids = processor.get_decoder_prompt_ids(language="ml", task="transcribe")
asr = pipeline(
        "automatic-speech-recognition", model="thennal/whisper-medium-ml", device=0,
    )
transcription = asr(audio, chunk_length_s=30, max_new_tokens=448, return_timestamps=False,  generate_kwargs={
        "forced_decoder_ids": forced_decoder_ids, 
        "do_sample": True,
    })
```
## Model description

More information needed

## Intended uses & limitations

More information needed

## Training and evaluation data

More information needed

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 32
- eval_batch_size: 16
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- training_steps: 8000
- mixed_precision_training: Native AMP

### Framework versions

- Transformers 4.26.0.dev0
- Pytorch 1.13.0+cu117
- Datasets 2.7.1.dev0
- Tokenizers 0.13.2