wav2vec2-xls-r-300m-bp1-es_eu Open-Source Model - Efficient Implementation of Basque Automatic Speech Recognition

Wav2vec2 Xls R 300m Bp1 Es Eu

Developed by gttsehu

A fine-tuned Basque automatic speech recognition model based on facebook/wav2vec2-xls-r-300m, achieving a 3.67% word error rate on the Basque Parliament dataset

Speech Recognition

Transformers

Open Source License:Apache-2.0 #Basque speech recognition #Parliament scenario optimization #Low word error rate

Downloads 49

Release Time : 1/5/2024

Model Overview

This is an automatic speech recognition model optimized for Basque parliamentary scenarios, based on the XLS-R architecture, particularly suitable for political speeches and meeting minutes in Basque.

Model Features

Low word error rate

Achieves only 3.67% word error rate on the Basque Parliament dataset, demonstrating excellent performance

Multi-GPU optimization

Utilizes 4-GPU distributed training, supporting efficient large-scale data processing

Domain adaptation

Specifically optimized for political speech scenarios, adapting to the special vocabulary and expressions in parliamentary meetings

Model Capabilities

Basque speech-to-text

Automatic meeting minutes generation

Political speech recognition

Use Cases

Government meeting records

Automatic parliamentary meeting transcription

Automatically transcribes Basque parliamentary meeting content to generate text records

Accuracy rate of 96.33% (WER 3.67%)

Speech archive digitization

Historical meeting audio transcription

Converts stored historical audio recordings of Basque Parliament into searchable text

🚀 facebook/wav2vec2-xls-r-300m

This is a fine - tuned model for automatic speech recognition, achieving excellent results on the evaluation set.

🚀 Quick Start

This work was partially funded by the Spanish Ministry of Science and Innovation (OPENSPEECH project, PID2019 - 106424RB - I00).

This model is a fine - tuned version of facebook/wav2vec2-xls-r-300m on the GTTSEHU/BASQUE_PARLIAMENT_1 - NA dataset. It achieves the following results on the evaluation set:

Loss: 0.0846
Wer: 0.0367
Cer: 0.0132

📚 Documentation

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0001
train_batch_size: 4
eval_batch_size: 4
seed: 42
distributed_type: multi - GPU
num_devices: 4
gradient_accumulation_steps: 2
total_train_batch_size: 32
total_eval_batch_size: 16
optimizer: Adam with betas=(0.9,0.999) and epsilon = 1e - 08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 1000
num_epochs: 6.0
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.7054	0.19	4000	0.1011	0.0871	0.0227
0.0856	0.39	8000	0.0995	0.0747	0.0207
0.075	0.58	12000	0.0868	0.0647	0.0185
0.0694	0.77	16000	0.0853	0.0619	0.0183
0.0658	0.97	20000	0.0778	0.0573	0.0171
0.0589	1.16	24000	0.0821	0.0546	0.0166
0.0572	1.35	28000	0.0827	0.0558	0.0170
0.0551	1.55	32000	0.0830	0.0533	0.0169
0.054	1.74	36000	0.0788	0.0512	0.0162
0.0524	1.93	40000	0.0783	0.0489	0.0156
0.048	2.13	44000	0.0861	0.0492	0.0160
0.046	2.32	48000	0.0763	0.0494	0.0154
0.0456	2.51	52000	0.0835	0.0471	0.0153
0.0439	2.71	56000	0.0790	0.0469	0.0152
0.0436	2.9	60000	0.0832	0.0472	0.0155
0.0406	3.09	64000	0.0810	0.0442	0.0148
0.0386	3.29	68000	0.0810	0.0436	0.0146
0.038	3.48	72000	0.0778	0.0430	0.0143
0.0373	3.67	76000	0.0785	0.0430	0.0144
0.0363	3.87	80000	0.0788	0.0421	0.0144
0.0348	4.06	84000	0.0823	0.0423	0.0144
0.0323	4.25	88000	0.0819	0.0407	0.0143
0.0319	4.45	92000	0.0809	0.0410	0.0142
0.0314	4.64	96000	0.0821	0.0400	0.0138
0.0306	4.83	100000	0.0813	0.0389	0.0137
0.0295	5.03	104000	0.0820	0.0377	0.0131
0.0275	5.22	108000	0.0866	0.0378	0.0137
0.0267	5.41	112000	0.0831	0.0376	0.0134
0.0264	5.61	116000	0.0845	0.0369	0.0132
0.0258	5.8	120000	0.0859	0.0370	0.0133
0.0254	6.0	124000	0.0846	0.0367	0.0132

Framework versions

Transformers 4.36.0.dev0
Pytorch 2.1.1+cu121
Datasets 2.15.1.dev0
Tokenizers 0.15.0

📄 License

apache - 2.0

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