metadata

language:
  - as
license: apache-2.0
tags:
  - automatic-speech-recognition
  - mozilla-foundation/common_voice_8_0
  - generated_from_trainer
  - as
  - robust-speech-event
  - model_for_talk
  - hf-asr-leaderboard
datasets:
  - mozilla-foundation/common_voice_8_0
model-index:
  - name: wav2vec2-large-xls-r-300m-as-g1
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice 8
          type: mozilla-foundation/common_voice_8_0
          args: as
        metrics:
          - name: Test WER
            type: wer
            value: 0.6540934419202743
          - name: Test CER
            type: cer
            value: 0.21454042646095625
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Robust Speech Event - Dev Data
          type: speech-recognition-community-v2/dev_data
          args: as
        metrics:
          - name: Test WER
            type: wer
            value: NA
          - name: Test CER
            type: cer
            value: NA

wav2vec2-large-xls-r-300m-as-g1

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - AS dataset. It achieves the following results on the evaluation set:

Loss: 1.3327
Wer: 0.5744

Evaluation Commands

To evaluate on mozilla-foundation/common_voice_8_0 with test split

python eval.py --model_id DrishtiSharma/wav2vec2-large-xls-r-300m-as-g1 --dataset mozilla-foundation/common_voice_8_0 --config as --split test --log_outputs

To evaluate on speech-recognition-community-v2/dev_data

Assamese language isn't available in speech-recognition-community-v2/dev_data

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0003
train_batch_size: 16
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 2
total_train_batch_size: 32
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 1000
num_epochs: 200
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
14.1958	5.26	100	7.1919	1.0
5.0035	10.51	200	3.9362	1.0
3.6193	15.77	300	3.4451	1.0
3.4852	21.05	400	3.3536	1.0
2.8489	26.31	500	1.6451	0.9100
0.9568	31.56	600	1.0514	0.7561
0.4865	36.82	700	1.0434	0.7184
0.322	42.1	800	1.0825	0.7210
0.2383	47.36	900	1.1304	0.6897
0.2136	52.62	1000	1.1150	0.6854
0.179	57.87	1100	1.2453	0.6875
0.1539	63.15	1200	1.2211	0.6704
0.1303	68.41	1300	1.2859	0.6747
0.1183	73.67	1400	1.2775	0.6721
0.0994	78.92	1500	1.2321	0.6404
0.0991	84.21	1600	1.2766	0.6524
0.0887	89.46	1700	1.3026	0.6344
0.0754	94.72	1800	1.3199	0.6704
0.0693	99.97	1900	1.3044	0.6361
0.0568	105.26	2000	1.3541	0.6254
0.0536	110.51	2100	1.3320	0.6249
0.0529	115.77	2200	1.3370	0.6271
0.048	121.05	2300	1.2757	0.6031
0.0419	126.31	2400	1.2661	0.6172
0.0349	131.56	2500	1.2897	0.6048
0.0309	136.82	2600	1.2688	0.5962
0.0278	142.1	2700	1.2885	0.5954
0.0254	147.36	2800	1.2988	0.5915
0.0223	152.62	2900	1.3153	0.5941
0.0216	157.87	3000	1.2936	0.5937
0.0186	163.15	3100	1.2906	0.5877
0.0156	168.41	3200	1.3476	0.5962
0.0158	173.67	3300	1.3363	0.5847
0.0142	178.92	3400	1.3367	0.5847
0.0153	184.21	3500	1.3105	0.5757
0.0119	189.46	3600	1.3255	0.5705
0.0115	194.72	3700	1.3340	0.5787
0.0103	199.97	3800	1.3327	0.5744

Framework versions

Transformers 4.16.2
Pytorch 1.10.0+cu111
Datasets 1.18.3
Tokenizers 0.11.0