wav2vec2-large-xls-r-300m-bulgarian Open Source Model - Accurately Realize Bulgarian Speech Recognition

Wav2vec2 Large Xls R 300m Bulgarian

Developed by infinitejoy

A Bulgarian speech recognition model fine-tuned on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - BG dataset based on facebook/wav2vec2-xls-r-300m

Speech Recognition

Transformers

OtherOpen Source License:Apache-2.0 #Bulgarian Speech Recognition #Multi-scenario Speech Transcription #Low-resource Language Optimization

Downloads 10.59k

Release Time : 3/2/2022

Model Overview

This is a model for Bulgarian Automatic Speech Recognition (ASR), based on the XLS-R architecture and fine-tuned on the Bulgarian dataset from Common Voice 7.0.

Model Features

Multilingual Pretraining

Fine-tuned based on the XLS-R-300M multilingual model with strong speech representation capabilities

Bulgarian Language Optimization

Specifically fine-tuned for Bulgarian to adapt to its unique linguistic features

Medium Scale

300M parameter size, balancing performance and resource consumption

Model Capabilities

Bulgarian Speech Recognition

Speech-to-Text Conversion

Conversation Transcription

Use Cases

Speech Transcription

Voice Memo Transcription

Convert Bulgarian voice memos into text

WER 46.68% on Common Voice 7 test set

Customer Service Dialogue Recording

Automatically transcribe Bulgarian customer service conversations

WER 64.08% on Robust Speech Event test data

Assistive Technology

Voice Control Applications

Provide voice control interfaces for Bulgarian users

🚀 XLS-R-300M - Bulgarian

This is a fine - tuned model for Automatic Speech Recognition on Bulgarian data, achieving specific performance metrics on evaluation sets.

🚀 Quick Start

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

Loss: 0.4487
Wer: 0.4674

📚 Documentation

Model Details

Property	Details
Model Type	Fine - tuned version of `facebook/wav2vec2-xls-r-300m` on `MOZILLA - FOUNDATION/COMMON_VOICE_7_0 - BG`
Training Datasets	`mozilla - foundation/common_voice_7_0`
Results	Task: Automatic Speech Recognition - Dataset: Common Voice 7 (bg): Test WER = 46.68, Test CER = 10.75 - Dataset: Robust Speech Event - Dev Data (bg): Test WER = 63.68, Test CER = 19.88 - Dataset: Robust Speech Event - Test Data (bg): Test WER = 64.08, Test CER = 19.88

Training Procedure

Training Hyperparameters

The following hyperparameters were used during training:

learning_rate: 7e - 05
train_batch_size: 32
eval_batch_size: 32
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon = 1e - 08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 500
num_epochs: 100.0
mixed_precision_training: Native AMP

Training Results

Training Loss	Epoch	Step	Validation Loss	Wer
2.9774	6.33	500	2.9769	1.0
1.3453	12.66	1000	0.6523	0.6980
1.1658	18.99	1500	0.5636	0.6359
1.0797	25.32	2000	0.5004	0.5759
1.044	31.65	2500	0.4958	0.5569
0.9915	37.97	3000	0.4971	0.5350
0.9429	44.3	3500	0.4829	0.5229
0.9266	50.63	4000	0.4515	0.5074
0.8965	56.96	4500	0.4599	0.5039
0.878	63.29	5000	0.4735	0.4954
0.8494	69.62	5500	0.4460	0.4878
0.8343	75.95	6000	0.4510	0.4795
0.8236	82.28	6500	0.4538	0.4789
0.8069	88.61	7000	0.4526	0.4748
0.7958	94.94	7500	0.4496	0.4700

Framework Versions

Transformers 4.16.0.dev0
Pytorch 1.10.1+cu102
Datasets 1.17.1.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