|
--- |
|
language: |
|
- hy |
|
license: apache-2.0 |
|
tags: |
|
- automatic-speech-recognition |
|
- mozilla-foundation/common_voice_8_0 |
|
- generated_from_trainer |
|
- robust-speech-event |
|
- hy |
|
- hf-asr-leaderboard |
|
datasets: |
|
- common_voice |
|
model-index: |
|
- name: wav2vec2-xls-r-300m-hy |
|
results: |
|
- task: |
|
type: automatic-speech-recognition |
|
name: Speech Recognition |
|
dataset: |
|
type: mozilla-foundation/common_voice_8_0 |
|
name: Common Voice hy-AM |
|
args: hy-AM |
|
metrics: |
|
- type: wer |
|
value: 13.192818110850899 |
|
name: WER LM |
|
- type: cer |
|
value: 2.787051087506323 |
|
name: CER LM |
|
- task: |
|
name: Speech Recognition |
|
type: automatic-speech-recognition |
|
dataset: |
|
name: Robust Speech Event - Dev Data |
|
type: speech-recognition-community-v2/dev_data |
|
args: hy |
|
metrics: |
|
- name: Test WER |
|
type: wer |
|
value: 22.246048764990867 |
|
- name: Test CER |
|
type: cer |
|
value: 7.59406739840239 |
|
--- |
|
|
|
<!-- 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. --> |
|
|
|
# |
|
|
|
This model is a fine-tuned version of [facebook/wav2vec2-xls-r-300m](https://huggingface.co/facebook/wav2vec2-xls-r-300m) on the /WORKSPACE/DATA/HY/NOIZY_STUDENT_3/ - NA dataset. |
|
It achieves the following results on the evaluation set: |
|
- Loss: 0.2293 |
|
- Wer: 0.3333 |
|
- Cer: 0.0602 |
|
|
|
## 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: 7e-05 |
|
- train_batch_size: 64 |
|
- eval_batch_size: 64 |
|
- seed: 842 |
|
- gradient_accumulation_steps: 2 |
|
- total_train_batch_size: 128 |
|
- optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-08 |
|
- lr_scheduler_type: cosine |
|
- lr_scheduler_warmup_ratio: 0.1 |
|
- training_steps: 4000 |
|
- mixed_precision_training: Native AMP |
|
|
|
### Training results |
|
|
|
| Training Loss | Epoch | Step | Validation Loss | Wer | Cer | |
|
|:-------------:|:-----:|:----:|:---------------:|:------:|:------:| |
|
| 3.1471 | 7.02 | 400 | 3.1599 | 1.0 | 1.0 | |
|
| 1.8691 | 14.04 | 800 | 0.7674 | 0.7361 | 0.1686 | |
|
| 1.3227 | 21.05 | 1200 | 0.3849 | 0.5336 | 0.1007 | |
|
| 1.163 | 28.07 | 1600 | 0.3015 | 0.4559 | 0.0823 | |
|
| 1.0768 | 35.09 | 2000 | 0.2721 | 0.4032 | 0.0728 | |
|
| 1.0224 | 42.11 | 2400 | 0.2586 | 0.3825 | 0.0691 | |
|
| 0.9817 | 49.12 | 2800 | 0.2458 | 0.3653 | 0.0653 | |
|
| 0.941 | 56.14 | 3200 | 0.2306 | 0.3388 | 0.0605 | |
|
| 0.9235 | 63.16 | 3600 | 0.2315 | 0.3380 | 0.0615 | |
|
| 0.9141 | 70.18 | 4000 | 0.2293 | 0.3333 | 0.0602 | |
|
|
|
|
|
### Framework versions |
|
|
|
- Transformers 4.17.0.dev0 |
|
- Pytorch 1.10.2 |
|
- Datasets 1.18.4.dev0 |
|
- Tokenizers 0.11.0 |
|
|