Hello_2b_3 Open-Source Automatic Speech Recognition Model - Free Deployment for Precise Recognition of Turkish Speech

Hello 2b 3

Developed by patrickvonplaten

This model is an automatic speech recognition (ASR) model fine-tuned on the Turkish (TR) COMMON_VOICE dataset based on facebook/wav2vec2-xls-r-2b

Speech Recognition

Transformers

Other#Turkish speech recognition #Multi-GPU fine-tuning #Low-resource optimization

Downloads 21

Release Time : 3/2/2022

Model Overview

For Turkish automatic speech recognition tasks, fine-tuned based on the wav2vec2-xls-r-2b architecture

Model Features

Large-scale pretrained model fine-tuning

Fine-tuned from the 2-billion-parameter wav2vec2-xls-r model with powerful speech feature extraction capabilities

Turkish language optimization

Specifically optimized for Turkish speech recognition tasks

Multi-GPU training

Distributed training using 2 GPUs to improve training efficiency

Model Capabilities

Turkish speech-to-text

Continuous speech recognition

Audio content transcription

Use Cases

Speech transcription

Turkish speech transcription

Convert Turkish speech content to text

Word Error Rate (WER) of 0.9808 on the validation set

Voice assistants

Turkish voice command recognition

For voice command recognition in Turkish voice assistants

🚀 hello_2b_3

This model is a fine - tuned version of [facebook/wav2vec2 - xls - r - 2b](https://huggingface.co/facebook/wav2vec2 - xls - r - 2b) on the COMMON_VOICE - TR dataset. It offers an effective solution for automatic speech recognition tasks, leveraging the pre - trained capabilities of the base model and adapting them to the Turkish language.

🚀 Quick Start

This model is a fine - tuned version of [facebook/wav2vec2 - xls - r - 2b](https://huggingface.co/facebook/wav2vec2 - xls - r - 2b) on the COMMON_VOICE - TR dataset. It achieves the following results on the evaluation set:

Loss: 1.5615
Wer: 0.9808

📚 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: 5e - 06
train_batch_size: 2
eval_batch_size: 8
seed: 42
distributed_type: multi - GPU
num_devices: 2
gradient_accumulation_steps: 8
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: 500
num_epochs: 30.0
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.6389	0.92	100	3.6218	1.0
1.6676	1.85	200	3.2655	1.0
0.3067	2.77	300	3.2273	1.0
0.1924	3.7	400	3.0238	0.9999
0.1777	4.63	500	2.1606	0.9991
0.1481	5.55	600	1.8742	0.9982
0.1128	6.48	700	2.0114	0.9994
0.1806	7.4	800	1.9032	0.9984
0.0399	8.33	900	2.0556	0.9996
0.0729	9.26	1000	2.0515	0.9987
0.0847	10.18	1100	2.2121	0.9995
0.0777	11.11	1200	1.7002	0.9923
0.0476	12.04	1300	1.5262	0.9792
0.0518	12.96	1400	1.5990	0.9832
0.071	13.88	1500	1.6326	0.9875
0.0333	14.81	1600	1.5955	0.9870
0.0369	15.74	1700	1.5577	0.9832
0.0689	16.66	1800	1.5415	0.9839
0.0227	17.59	1900	1.5450	0.9878
0.0472	18.51	2000	1.5642	0.9846
0.0214	19.44	2100	1.6103	0.9846
0.0289	20.37	2200	1.6467	0.9898
0.0182	21.29	2300	1.5268	0.9780
0.0439	22.22	2400	1.6001	0.9818
0.06	23.15	2500	1.5481	0.9813
0.0351	24.07	2600	1.5672	0.9820
0.0198	24.99	2700	1.6303	0.9856
0.0328	25.92	2800	1.5958	0.9831
0.0245	26.85	2900	1.5745	0.9809
0.0885	27.77	3000	1.5455	0.9809
0.0224	28.7	3100	1.5378	0.9824
0.0223	29.63	3200	1.5642	0.9810

Framework versions

Transformers 4.13.0.dev0
Pytorch 1.10.0
Datasets 1.15.2.dev0
Tokenizers 0.10.3

Property	Details
Model Type	Fine - tuned version of facebook/wav2vec2 - xls - r - 2b on COMMON_VOICE - TR dataset
Training Data	COMMON_VOICE - TR dataset

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