w11wo/wav2vec2-xls-r-300m-zh-HK-lm-v2

Wav2Vec2 XLS-R 300M Cantonese (zh-HK) LM

Wav2Vec2 XLS-R 300M Cantonese (zh-HK) 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 zh-HK subset of the Common Voice dataset. A 5-gram Language model, trained on multiple PyCantonese corpora, 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-zh-HK-lm-v2	300M	XLS-R	Common Voice zh-HK Dataset

Evaluation Results

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

Dataset	CER
Common Voice	31.73%
Common Voice 7	23.11%
Common Voice 8	23.02%
Robust Speech Event - Dev Data	56.60%

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

Dataset	CER
Common Voice	24.09%
Common Voice 7	23.10%
Common Voice 8	23.02%
Robust Speech Event - Dev Data	56.86%

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: 0.0001
• 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: 100.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
69.8341	1.34	500	80.0722	1.0	1.0
6.6418	2.68	1000	6.6346	1.0	1.0
6.2419	4.02	1500	6.2909	1.0	1.0
6.0813	5.36	2000	6.1150	1.0	1.0
5.9677	6.7	2500	6.0301	1.1386	1.0028
5.9296	8.04	3000	5.8975	1.2113	1.0058
5.6434	9.38	3500	5.5404	2.1624	1.0171
5.1974	10.72	4000	4.5440	2.1702	0.9366
4.3601	12.06	4500	3.3839	2.2464	0.8998
3.9321	13.4	5000	2.8785	2.3097	0.8400
3.6462	14.74	5500	2.5108	1.9623	0.6663
3.5156	16.09	6000	2.2790	1.6479	0.5706
3.32	17.43	6500	2.1450	1.8337	0.6244
3.1918	18.77	7000	1.8536	1.9394	0.6017
3.1139	20.11	7500	1.7205	1.9112	0.5638
2.8995	21.45	8000	1.5478	1.0624	0.3250
2.7572	22.79	8500	1.4068	1.1412	0.3367
2.6881	24.13	9000	1.3312	2.0100	0.5683
2.5993	25.47	9500	1.2553	2.0039	0.6450
2.5304	26.81	10000	1.2422	2.0394	0.5789
2.4352	28.15	10500	1.1582	1.9970	0.5507
2.3795	29.49	11000	1.1160	1.8255	0.4844
2.3287	30.83	11500	1.0775	1.4123	0.3780
2.2622	32.17	12000	1.0704	1.7445	0.4894
2.2225	33.51	12500	1.0272	1.7237	0.5058
2.1843	34.85	13000	0.9756	1.8042	0.5028
2.1	36.19	13500	0.9527	1.8909	0.6055
2.0741	37.53	14000	0.9418	1.9026	0.5880
2.0179	38.87	14500	0.9363	1.7977	0.5246
2.0615	40.21	15000	0.9635	1.8112	0.5599
1.9448	41.55	15500	0.9249	1.7250	0.4914
1.8966	42.89	16000	0.9023	1.5829	0.4319
1.8662	44.24	16500	0.9002	1.4833	0.4230
1.8136	45.58	17000	0.9076	1.1828	0.2987
1.7908	46.92	17500	0.8774	1.5773	0.4258
1.7354	48.26	18000	0.8727	1.5037	0.4024
1.6739	49.6	18500	0.8636	1.1239	0.2789
1.6457	50.94	19000	0.8516	1.2269	0.3104
1.5847	52.28	19500	0.8399	1.3309	0.3360
1.5971	53.62	20000	0.8441	1.3153	0.3335
1.602	54.96	20500	0.8590	1.2932	0.3433
1.5063	56.3	21000	0.8334	1.1312	0.2875
1.4631	57.64	21500	0.8474	1.1698	0.2999
1.4997	58.98	22000	0.8638	1.4279	0.3854
1.4301	60.32	22500	0.8550	1.2737	0.3300
1.3798	61.66	23000	0.8266	1.1802	0.2934
1.3454	63.0	23500	0.8235	1.3816	0.3711
1.3678	64.34	24000	0.8550	1.6427	0.5035
1.3761	65.68	24500	0.8510	1.6709	0.4907
1.2668	67.02	25000	0.8515	1.5842	0.4505
1.2835	68.36	25500	0.8283	1.5353	0.4221
1.2961	69.7	26000	0.8339	1.5743	0.4369
1.2656	71.05	26500	0.8331	1.5331	0.4217
1.2556	72.39	27000	0.8242	1.4708	0.4109
1.2043	73.73	27500	0.8245	1.4469	0.4031
1.2722	75.07	28000	0.8202	1.4924	0.4096
1.202	76.41	28500	0.8290	1.3807	0.3719
1.1679	77.75	29000	0.8195	1.4097	0.3749
1.1967	79.09	29500	0.8059	1.2074	0.3077
1.1241	80.43	30000	0.8137	1.2451	0.3270
1.1414	81.77	30500	0.8117	1.2031	0.3121
1.132	83.11	31000	0.8234	1.4266	0.3901
1.0982	84.45	31500	0.8064	1.3712	0.3607
1.0797	85.79	32000	0.8167	1.3356	0.3562
1.0119	87.13	32500	0.8215	1.2754	0.3268
1.0216	88.47	33000	0.8163	1.2512	0.3184
1.0375	89.81	33500	0.8137	1.2685	0.3290
0.9794	91.15	34000	0.8220	1.2724	0.3255
1.0207	92.49	34500	0.8165	1.2906	0.3361
1.0169	93.83	35000	0.8153	1.2819	0.3305
1.0127	95.17	35500	0.8187	1.2832	0.3252
0.9978	96.51	36000	0.8111	1.2612	0.3210
0.9923	97.85	36500	0.8076	1.2278	0.3122
1.0451	99.2	37000	0.8086	1.2451	0.3156

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 Cantonese (zh-HK) 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 1.18.4.dev0
• Tokenizers 0.11.0