Wav2Vec2 Open-Source Russian Speech Recognition Model - Free to Use, Accurately Recognize Speech with Low Word Error Rate

Wav2vec2 Large Xls R 300m Russian Colab Beam Search Test

Developed by jfealko

This model is a Russian speech recognition model fine-tuned on the Common Voice dataset based on facebook/wav2vec2-xls-r-300m, achieving a word error rate of 0.468 on the evaluation set.

Speech Recognition

Transformers

Open Source License:Apache-2.0 #Russian Speech Recognition #Large Model Fine-tuning #Low Word Error Rate

Downloads 18

Release Time : 4/7/2022

Model Overview

This is a speech recognition model optimized for Russian, fine-tuned based on the wav2vec2-xls-r-300m architecture, suitable for Russian speech-to-text tasks.

Model Features

Russian Optimization

Specifically fine-tuned for Russian speech data, improving the accuracy of Russian recognition.

Low Word Error Rate

Achieved a word error rate of 0.468 on the evaluation set, demonstrating good performance.

Fine-tuned from Large Model

Fine-tuned from the facebook/wav2vec2-xls-r-300m large model, inheriting its powerful speech feature extraction capabilities.

Model Capabilities

Russian Speech Recognition

Speech-to-Text

Automatic Speech Recognition

Use Cases

Speech Transcription

Russian Meeting Minutes

Automatically transcribe Russian meeting recordings into text

Accuracy approximately 53.2% (word error rate 0.468)

Russian Voice Assistant

Used as the speech recognition module for Russian voice assistants

Education

Russian Learning Aid

Assist Russian learners in checking pronunciation accuracy

🚀 wav2vec2-large-xls-r-300m-russian-colab-beam_search_test

This model is a fine - tuned version of facebook/wav2vec2-xls-r-300m on the common_voice dataset. It offers significant value in speech - related tasks by leveraging the pre - trained knowledge and fine - tuning on specific data.

🚀 Quick Start

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

Loss: 0.7619
Wer: 0.4680

📚 Documentation

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: 0.0003
train_batch_size: 16
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 64
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e - 08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 800
num_epochs: 100
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
10.0158	4.16	100	5.4134	1.0
4.0394	8.33	200	3.4304	1.0
3.2721	12.49	300	3.2273	1.0
3.1277	16.66	400	2.8023	0.9984
1.3791	20.82	500	0.9888	0.8546
0.3659	24.99	600	0.7602	0.6304
0.1858	29.16	700	0.7965	0.6156
0.1403	33.33	800	0.7998	0.5839
0.1173	37.49	900	0.8353	0.5941
0.0917	41.66	1000	0.8272	0.5522
0.0743	45.82	1100	0.8342	0.5471
0.063	49.99	1200	0.7988	0.5352
0.0528	54.16	1300	0.7740	0.5201
0.0456	58.33	1400	0.7636	0.5165
0.0389	62.49	1500	0.7922	0.5161
0.0329	66.66	1600	0.8035	0.5158
0.0283	70.82	1700	0.7873	0.4832
0.0255	74.99	1800	0.7853	0.4870
0.0236	79.16	1900	0.8236	0.5045
0.0202	83.33	2000	0.7661	0.4796
0.0165	87.49	2100	0.7584	0.4680
0.0156	91.66	2200	0.7685	0.4772
0.0149	95.82	2300	0.7519	0.4696
0.0126	99.99	2400	0.7619	0.4680

Framework versions

Transformers 4.11.3
Pytorch 1.10.0+cu111
Datasets 1.18.3
Tokenizers 0.10.3

📄 License

This project is licensed under the Apache - 2.0 license.

Featured Recommended AI Models

Empowering the Future, Your AI Solution Knowledge Base

English 简体中文繁體中文にほんご