metadata

language:
  - hu
license: apache-2.0
tags:
  - automatic-speech-recognition
  - mozilla-foundation/common_voice_7_0
  - generated_from_trainer
  - hu
  - robust-speech-event
  - model_for_talk
datasets:
  - mozilla-foundation/common_voice_7_0
model-index:
  - name: Akashpb13/xlsr_hungarian_new
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice 7
          type: mozilla-foundation/common_voice_7_0
          args: hu
        metrics:
          - name: Test WER
            type: wer
            value: 0.02698525418772714
          - name: Test CER
            type: cer
            value: 0.005033063261641211
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Robust Speech Event - Dev Data
          type: speech-recognition-community-v2/dev_data
          args: hu
        metrics:
          - name: Test WER
            type: wer
            value: 0.02698525418772714
          - name: Test CER
            type: cer
            value: 0.005033063261641211

Akashpb13/xlsr_hungarian_new

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - hu dataset. It achieves the following results on evaluation set (which is 10 percent of train data set merged with invalidated data, reported, other, dev and validated datasets):

Loss: 0.184265
Wer: 0.292771

Model description

"facebook/wav2vec2-xls-r-300m" was finetuned.

Intended uses & limitations

More information needed

Training and evaluation data

Training data - Common voice hungarian train.tsv, dev.tsv, invalidated.tsv, reported.tsv, other.tsv and validated.tsv Only those points were considered where upvotes were greater than downvotes and duplicates were removed after concatenation of all the datasets given in common voice 7.0

Training procedure

For creating the train dataset, all possible datasets were appended and 90-10 split was used.

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.000095637994662983496
train_batch_size: 16
eval_batch_size: 16
seed: 13
gradient_accumulation_steps: 16
total_train_batch_size: 316
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine_with_restarts
lr_scheduler_warmup_steps: 500
num_epochs: 100
mixed_precision_training: Native AMP

Training results

Step	Training Loss	Validation Loss	Wer
500	4.825900	1.001413	0.810308
1000	0.561400	0.202275	0.361987
1500	0.298900	0.169643	0.326449
2000	0.236500	0.168602	0.316215
2500	0.199100	0.182484	0.308587
3000	0.179100	0.178076	0.303005
3500	0.161500	0.179107	0.299935
4000	0.151700	0.183371	0.295283
4500	0.143700	0.184443	0.295283
5000	0.138900	0.184265	0.292771

Framework versions

Transformers 4.16.0.dev0
Pytorch 1.10.0+cu102
Datasets 1.17.1.dev0
Tokenizers 0.10.3

Evaluation Commands

To evaluate on mozilla-foundation/common_voice_7_0 with split test

python eval.py --model_id Akashpb13/xlsr_hungarian_new --dataset mozilla-foundation/common_voice_7_0 --config hu --split test