metadata

language: ko
license: apache-2.0
tags:
  - automatic-speech-recognition
  - generated_from_trainer
  - hf-asr-leaderboard
  - robust-speech-event
datasets:
  - kresnik/zeroth_korean
model-index:
  - name: Wav2Vec2 XLS-R 300M Korean LM
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Zeroth Korean
          type: kresnik/zeroth_korean
          args: clean
        metrics:
          - name: Test WER
            type: wer
            value: 30.94
          - name: Test CER
            type: cer
            value: 7.97
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Robust Speech Event - Dev Data
          type: speech-recognition-community-v2/dev_data
          args: ko
        metrics:
          - name: Test WER
            type: wer
            value: 68.34
          - name: Test CER
            type: cer
            value: 37.08
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Robust Speech Event - Test Data
          type: speech-recognition-community-v2/eval_data
          args: ko
        metrics:
          - name: Test WER
            type: wer
            value: 66.47

Wav2Vec2 XLS-R 300M Korean LM

Wav2Vec2 XLS-R 300M Korean LM is an automatic speech recognition model based on the XLS-R architecture. This model is a fine-tuned version of Wav2Vec2-XLS-R-300M on the Zeroth Korean dataset. A 5-gram Language model, trained on the Korean subset of Open Subtitles, was then subsequently added to this model.

This model was trained using HuggingFace's PyTorch framework and is part of the Robust Speech Challenge Event organized by HuggingFace. All training was done on a Tesla V100, sponsored by OVH.

All necessary scripts used for training could be found in the Files and versions tab, as well as the Training metrics logged via Tensorboard.

As for the N-gram language model training, we followed the blog post tutorial provided by HuggingFace.

Model

Model	#params	Arch.	Training/Validation data (text)
`wav2vec2-xls-r-300m-korean-lm`	300M	XLS-R	`Zeroth Korean` Dataset

Evaluation Results

The model achieves the following results on evaluation without a language model:

Dataset	WER	CER
`Zeroth Korean`	29.54%	9.53%
`Robust Speech Event - Dev Data`	76.26%	38.67%

With the addition of the language model, it achieves the following results:

Dataset	WER	CER
`Zeroth Korean`	30.94%	7.97%
`Robust Speech Event - Dev Data`	68.34%	37.08%

Training procedure

The training process did not involve the addition of a language model. The following results were simply lifted from the original automatic speech recognition model training.

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 7.5e-05
train_batch_size: 8
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 4
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: 2000
num_epochs: 50.0
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
19.7138	0.72	500	19.6427	1.0	1.0
4.8039	1.44	1000	4.7842	1.0	1.0
4.5619	2.16	1500	4.5608	0.9992	0.9598
4.254	2.88	2000	4.2729	0.9955	0.9063
4.1905	3.6	2500	4.2257	0.9903	0.8758
4.0683	4.32	3000	3.9294	0.9937	0.7911
3.486	5.04	3500	2.7045	1.0012	0.5934
2.946	5.75	4000	1.9691	0.9425	0.4634
2.634	6.47	4500	1.5212	0.8807	0.3850
2.4066	7.19	5000	1.2551	0.8177	0.3601
2.2651	7.91	5500	1.0423	0.7650	0.3039
2.1828	8.63	6000	0.9599	0.7273	0.3106
2.1023	9.35	6500	0.9482	0.7161	0.3063
2.0536	10.07	7000	0.8242	0.6767	0.2860
1.9803	10.79	7500	0.7643	0.6563	0.2637
1.9468	11.51	8000	0.7319	0.6441	0.2505
1.9178	12.23	8500	0.6937	0.6320	0.2489
1.8515	12.95	9000	0.6443	0.6053	0.2196
1.8083	13.67	9500	0.6286	0.6122	0.2148
1.819	14.39	10000	0.6015	0.5986	0.2074
1.7684	15.11	10500	0.5682	0.5741	0.1982
1.7195	15.83	11000	0.5385	0.5592	0.2007
1.7044	16.55	11500	0.5362	0.5524	0.2097
1.6879	17.27	12000	0.5119	0.5489	0.2083
1.656	17.98	12500	0.4990	0.5362	0.1968
1.6122	18.7	13000	0.4561	0.5092	0.1900
1.5919	19.42	13500	0.4778	0.5225	0.1975
1.5896	20.14	14000	0.4563	0.5098	0.1859
1.5589	20.86	14500	0.4362	0.4940	0.1725
1.5353	21.58	15000	0.4140	0.4826	0.1580
1.5441	22.3	15500	0.4031	0.4742	0.1550
1.5116	23.02	16000	0.3916	0.4748	0.1545
1.4731	23.74	16500	0.3841	0.4810	0.1542
1.4647	24.46	17000	0.3752	0.4524	0.1475
1.4328	25.18	17500	0.3587	0.4476	0.1461
1.4129	25.9	18000	0.3429	0.4242	0.1366
1.4062	26.62	18500	0.3450	0.4251	0.1355
1.3928	27.34	19000	0.3297	0.4145	0.1322
1.3906	28.06	19500	0.3210	0.4185	0.1336
1.358	28.78	20000	0.3131	0.3970	0.1275
1.3445	29.5	20500	0.3069	0.3920	0.1276
1.3159	30.22	21000	0.3035	0.3961	0.1255
1.3044	30.93	21500	0.2952	0.3854	0.1242
1.3034	31.65	22000	0.2966	0.3772	0.1227
1.2963	32.37	22500	0.2844	0.3706	0.1208
1.2765	33.09	23000	0.2841	0.3567	0.1173
1.2438	33.81	23500	0.2734	0.3552	0.1137
1.2487	34.53	24000	0.2703	0.3502	0.1118
1.2249	35.25	24500	0.2650	0.3484	0.1142
1.2229	35.97	25000	0.2584	0.3374	0.1097
1.2374	36.69	25500	0.2568	0.3337	0.1095
1.2153	37.41	26000	0.2494	0.3327	0.1071
1.1925	38.13	26500	0.2518	0.3366	0.1077
1.1908	38.85	27000	0.2437	0.3272	0.1057
1.1858	39.57	27500	0.2396	0.3265	0.1044
1.1808	40.29	28000	0.2373	0.3156	0.1028
1.1842	41.01	28500	0.2356	0.3152	0.1026
1.1668	41.73	29000	0.2319	0.3188	0.1025
1.1448	42.45	29500	0.2293	0.3099	0.0995
1.1327	43.17	30000	0.2265	0.3047	0.0979
1.1307	43.88	30500	0.2222	0.3078	0.0989
1.1419	44.6	31000	0.2215	0.3038	0.0981
1.1231	45.32	31500	0.2193	0.3013	0.0972
1.139	46.04	32000	0.2162	0.3007	0.0968
1.1114	46.76	32500	0.2122	0.2982	0.0960
1.111	47.48	33000	0.2125	0.2946	0.0948
1.0982	48.2	33500	0.2099	0.2957	0.0953
1.109	48.92	34000	0.2092	0.2955	0.0955
1.0905	49.64	34500	0.2088	0.2954	0.0953

Disclaimer

Do consider the biases which came from pre-training datasets that may be carried over into the results of this model.

Authors

Wav2Vec2 XLS-R 300M Korean LM was trained and evaluated by Wilson Wongso. All computation and development are done on OVH Cloud.

Framework versions

Transformers 4.17.0.dev0
Pytorch 1.10.2+cu102
Datasets 1.18.2.dev0
Tokenizers 0.10.3