Hello_2b Open-Source Automatic Speech Recognition Model - Free Deployment for Accurate Recognition of Turkish Speech

Hello 2b

Developed by patrickvonplaten

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

Speech Recognition

Transformers

Other#Turkish speech recognition #Large model fine-tuning #Low-resource optimization

Downloads 27

Release Time : 3/2/2022

Model Overview

A fine-tuned model for Turkish speech recognition based on the wav2vec2-xls-r-2b architecture

Model Features

Large-scale pre-training foundation

Fine-tuned based on facebook's wav2vec2-xls-r-2b model, with powerful speech feature extraction capabilities

Turkish language optimization

Specially fine-tuned for Turkish, adapting to the phonetic characteristics of Turkish

Multi-GPU training

Utilizes multi-GPU distributed training to improve training efficiency

Model Capabilities

Turkish speech-to-text

Continuous speech recognition

Large-scale speech data processing

Use Cases

Speech transcription

Turkish speech transcription

Convert Turkish speech content into text

Achieved a word error rate of 0.9531 on the COMMON_VOICE-TR test set

Voice assistants

Turkish voice command recognition

Used as the front-end speech recognition module for Turkish voice assistants

🚀 hello_2b

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 is designed for automatic speech recognition, leveraging the Common Voice dataset. On the evaluation set, it achieves the following results:

Loss: 1.2725
Wer: 0.9531

📚 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: 1e - 05
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
3.1646	0.92	100	3.2106	1.0
0.368	1.85	200	2.9963	1.0
0.2252	2.77	300	2.8078	0.9999
0.1546	3.7	400	2.3458	0.9996
0.1468	4.63	500	2.0086	0.9986
0.1261	5.55	600	1.8269	0.9985
0.1206	6.48	700	1.7347	0.9956
0.1959	7.4	800	1.6819	0.9955
0.0502	8.33	900	1.6809	0.9965
0.0811	9.26	1000	1.6674	0.9916
0.0534	10.18	1100	1.5719	0.9898
0.0402	11.11	1200	1.4620	0.9821
0.057	12.04	1300	1.3015	0.9554
0.0385	12.96	1400	1.3798	0.9600
0.0422	13.88	1500	1.3538	0.9699
0.014	14.81	1600	1.2507	0.9443
0.0232	15.74	1700	1.3318	0.9465
0.0554	16.66	1800	1.2784	0.9462
0.0316	17.59	1900	1.2503	0.9481
0.0524	18.51	2000	1.3920	0.9604
0.0142	19.44	2100	1.4224	0.9698
0.0288	20.37	2200	1.3475	0.9635
0.0106	21.29	2300	1.2232	0.9264
0.0396	22.22	2400	1.3323	0.9615
0.0349	23.15	2500	1.2741	0.9587
0.0121	24.07	2600	1.2671	0.9586
0.0224	24.99	2700	1.3001	0.9611
0.0449	25.92	2800	1.2777	0.9572
0.0186	26.85	2900	1.2766	0.9607
0.0365	27.77	3000	1.2935	0.9598
0.0105	28.7	3100	1.2761	0.9588
0.021	29.63	3200	1.2686	0.9528

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