XLS-R-ET Open-Source Estonian Speech Recognition Model - Free Deployment for Accurate Speech Content Recognition

Xls R Et

Developed by shpotes

Estonian automatic speech recognition model fine-tuned on the wav2vec2-xls-r-300m architecture, trained on the Common Voice 7.0 dataset

Speech Recognition

Transformers

OtherOpen Source License:Apache-2.0 #Estonian ASR #Low Character Error Rate #Common Voice Training

Downloads 23

Release Time : 3/2/2022

Model Overview

This model is an automatic speech recognition (ASR) system optimized for Estonian, capable of converting Estonian audio into text

Model Features

Multi-dataset Validation

Comprehensively evaluated on Common Voice 7.0 and Robust Speech Challenge datasets

Efficient Training

Optimized training process using mixed-precision training and cosine learning rate scheduling

Low Character Error Rate

Achieves 7.54% character error rate (CER) on test sets

Model Capabilities

Estonian audio-to-text conversion

Speech recognition

Speech transcription

Use Cases

Speech Transcription

Voice Memo Transcription

Convert Estonian voice memos into searchable text

Achieves 34.75% word error rate on Common Voice test set

Assistive Technology

Voice Input System

Provide voice input support for Estonian users

🚀 Automatic Speech Recognition Model

This model is a fine - tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA - FOUNDATION/COMMON_VOICE_7_0 - ET dataset. It offers a solution for automatic speech recognition tasks, achieving certain performance metrics on the evaluation set.

🚀 Quick Start

This model is a fine - tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA - FOUNDATION/COMMON_VOICE_7_0 - ET dataset. It achieves the following results on the evaluation set:

Loss: 0.4835
Wer: 0.3475

✨ Features

Fine - tuned: Based on the facebook/wav2vec2-xls-r-300m model, fine - tuned on the MOZILLA - FOUNDATION/COMMON_VOICE_7_0 - ET dataset.
Multiple Metrics: Evaluated on multiple datasets with different metrics such as WER and CER.

📚 Documentation

Model Information

Property	Details
Model Type	Fine - tuned version of facebook/wav2vec2-xls-r-300m
Training Data	MOZILLA - FOUNDATION/COMMON_VOICE_7_0 - ET dataset

Model Performance

The model was evaluated on the following datasets:

Common Voice 7:
- Test WER: 0.34753420299077314
- Test CER: 0.07542956089330906
Robust Speech Event - Dev Data:
- Test WER: 47.17
Robust Speech Event - Test Data:
- Test WER: 54.72

Training Procedure

Training Hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0003
train_batch_size: 72
eval_batch_size: 72
seed: 42
gradient_accumulation_steps: 2
total_train_batch_size: 144
optimizer: Adam with betas=(0.9,0.999) and epsilon = 1e - 08
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 500
num_epochs: 100.0
mixed_precision_training: Native AMP

Training Results

Training Loss	Epoch	Step	Validation Loss	Wer
0.3825	12.5	500	0.4022	0.5059
0.1592	25.0	1000	0.4585	0.4456
0.1215	37.5	1500	0.4550	0.4164
0.0972	50.0	2000	0.4725	0.4088
0.0731	62.5	2500	0.4568	0.3824
0.0527	75.0	3000	0.4712	0.3653
0.0428	87.5	3500	0.4813	0.3520
0.0383	100.0	4000	0.4835	0.3475

Framework Versions

Transformers 4.16.0.dev0
Pytorch 1.10.1+cu102
Datasets 1.18.4.dev0
Tokenizers 0.11.0

📄 License

This model 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