anuragshas/wav2vec2-xls-r-300m-lv-cv8-with-lm

XLS-R-300M - Latvian

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

• Loss: 0.1660
• Wer: 0.1705

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: 7.5e-05
• train_batch_size: 32
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 50.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.489	2.56	400	3.3590	1.0
2.9903	5.13	800	2.9704	1.0001
1.6712	7.69	1200	0.6179	0.6566
1.2635	10.26	1600	0.3176	0.4531
1.0819	12.82	2000	0.2517	0.3508
1.0136	15.38	2400	0.2257	0.3124
0.9625	17.95	2800	0.1975	0.2311
0.901	20.51	3200	0.1986	0.2097
0.8842	23.08	3600	0.1904	0.2039
0.8542	25.64	4000	0.1847	0.1981
0.8244	28.21	4400	0.1805	0.1847
0.7689	30.77	4800	0.1736	0.1832
0.7825	33.33	5200	0.1698	0.1821
0.7817	35.9	5600	0.1758	0.1803
0.7488	38.46	6000	0.1663	0.1760
0.7171	41.03	6400	0.1636	0.1721
0.7222	43.59	6800	0.1663	0.1729
0.7156	46.15	7200	0.1633	0.1715
0.7121	48.72	7600	0.1666	0.1718

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2+cu102
• Datasets 1.18.2.dev0
• Tokenizers 0.11.0

Evaluation Commands

1. To evaluate on mozilla-foundation/common_voice_8_0 with split test

python eval.py --model_id anuragshas/wav2vec2-xls-r-300m-lv-cv8-with-lm --dataset mozilla-foundation/common_voice_8_0 --config lv --split test

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

python eval.py --model_id anuragshas/wav2vec2-xls-r-300m-lv-cv8-with-lm --dataset speech-recognition-community-v2/dev_data --config lv --split validation --chunk_length_s 5.0 --stride_length_s 1.0

Inference With LM

import torch
from datasets import load_dataset
from transformers import AutoModelForCTC, AutoProcessor
import torchaudio.functional as F
model_id = "anuragshas/wav2vec2-xls-r-300m-lv-cv8-with-lm"
sample_iter = iter(load_dataset("mozilla-foundation/common_voice_8_0", "lv", split="test", streaming=True, use_auth_token=True))
sample = next(sample_iter)
resampled_audio = F.resample(torch.tensor(sample["audio"]["array"]), 48_000, 16_000).numpy()
model = AutoModelForCTC.from_pretrained(model_id)
processor = AutoProcessor.from_pretrained(model_id)
input_values = processor(resampled_audio, return_tensors="pt").input_values
with torch.no_grad():
    logits = model(input_values).logits
transcription = processor.batch_decode(logits.numpy()).text
# => "domāju ka viņam viss labi"

Eval results on Common Voice 8 "test" (WER):

Without LM	With LM (run ./eval.py)
16.997	9.633