The open-source Hausa speech recognition model w2v-bert-2.0-naijavoices-clearglobal-hausa-500hr-v0 offers accurate recognition with a low error rate.

W2v Bert 2.0 Naijavoices Clearglobal Hausa 500hr V0

Developed by asr-africa

A Hausa speech recognition model fine-tuned from facebook/w2v-bert-2.0, trained on 500 hours of Hausa data with a word error rate of 7.47%

Speech Recognition

Transformers

Open Source License:MIT #Hausa speech recognition #Low word error rate #High-precision fine-tuning

Downloads 16

Release Time : 3/29/2025

Model Overview

This is a speech recognition model optimized for Hausa, fine-tuned based on Facebook's wav2vec 2.0 architecture, suitable for Hausa speech-to-text tasks.

Model Features

Low word error rate

Achieves a word error rate (WER) of 7.47% and a character error rate (CER) of 1.86% on the evaluation set

Hausa optimization

Specifically optimized for the phonetic characteristics of Hausa

Efficient training

Uses mixed-precision training and gradient accumulation techniques to optimize training efficiency

Model Capabilities

Hausa speech recognition

Speech-to-text

Audio content transcription

Use Cases

Speech transcription

Hausa meeting minutes

Automatically transcribe Hausa meeting recordings into text records

Accuracy approximately 92.53% (1-WER)

Voice assistant

Provide voice interaction capabilities for Hausa-speaking users

Education

Language learning tool

Help learners practice Hausa pronunciation and listening

🚀 w2v-bert-2.0-naijavoices-clearglobal-hausa-500hr-v0

This model is a fine - tuned version of the transformers library. It is a fine - tuned variant of facebook/w2v-bert-2.0 on the None dataset, offering enhanced performance in relevant tasks.

🚀 Quick Start

This section will guide you through the basic information and performance metrics of the w2v-bert-2.0-naijavoices-clearglobal-hausa-500hr-v0 model.

Model Information

Property	Details
Library Name	transformers
License	MIT
Base Model	facebook/w2v-bert-2.0
Tags	generated_from_trainer
Metrics	wer

Evaluation Results

This model achieves the following results on the evaluation set:

Loss: 0.1944
Wer: 0.0747
Cer: 0.0186

📚 Documentation

Model description

This model is a fine - tuned version of facebook/w2v-bert-2.0 on the None dataset.

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

🔧 Technical Details

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 9e - 05
train_batch_size: 64
eval_batch_size: 32
seed: 42
gradient_accumulation_steps: 2
total_train_batch_size: 128
optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon = 1e - 08 and optimizer_args = No additional optimizer arguments
lr_scheduler_type: linear
lr_scheduler_warmup_ratio: 0.025
num_epochs: 100.0
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.389	1.0	2396	0.1232	0.1422	0.0334
0.1003	2.0	4792	0.1169	0.1378	0.0314
0.0975	3.0	7188	0.1177	0.1355	0.0319
0.0881	4.0	9584	0.1106	0.1231	0.0294
0.0831	5.0	11980	0.1112	0.1246	0.0302
0.079	6.0	14376	0.1123	0.1227	0.0297
0.0756	7.0	16772	0.1014	0.1160	0.0273
0.0735	8.0	19168	0.1042	0.1146	0.0273
0.0703	9.0	21564	0.0968	0.1096	0.0261
0.0667	10.0	23960	0.0967	0.1106	0.0260
0.0652	11.0	26356	0.1012	0.1121	0.0266
0.0613	12.0	28752	0.1010	0.1111	0.0266
0.0603	13.0	31148	0.1001	0.1109	0.0262
0.0575	14.0	33544	0.0937	0.1066	0.0249
0.0539	15.0	35940	0.0957	0.1079	0.0254
0.0542	16.0	38336	0.0993	0.1048	0.0250
0.0494	17.0	40732	0.0956	0.1023	0.0244
0.0471	18.0	43128	0.0995	0.1034	0.0243
0.0434	19.0	45524	0.0950	0.1025	0.0241
0.0411	20.0	47920	0.0992	0.1006	0.0239
0.0376	21.0	50316	0.1023	0.1017	0.0242
0.0339	22.0	52712	0.1015	0.0991	0.0234
0.0315	23.0	55108	0.1108	0.1005	0.0243
0.0281	24.0	57504	0.1100	0.0970	0.0230
0.0243	25.0	59900	0.1088	0.0937	0.0224
0.0214	26.0	62296	0.1162	0.0929	0.0222
0.0196	27.0	64692	0.1173	0.0920	0.0221
0.0173	28.0	67088	0.1262	0.0928	0.0225
0.0145	29.0	69484	0.1252	0.0913	0.0222
0.0132	30.0	71880	0.1318	0.0877	0.0214
0.0115	31.0	74276	0.1342	0.0868	0.0211
0.0114	32.0	76672	0.1326	0.0884	0.0216
0.0098	33.0	79068	0.1352	0.0853	0.0210
0.0088	34.0	81464	0.1398	0.0856	0.0209
0.008	35.0	83860	0.1432	0.0860	0.0210
0.0072	36.0	86256	0.1475	0.0840	0.0207
0.0065	37.0	88652	0.1444	0.0836	0.0206
0.006	38.0	91048	0.1467	0.0845	0.0208
0.006	39.0	93444	0.1501	0.0854	0.0210
0.0055	40.0	95840	0.1545	0.0841	0.0209
0.0047	41.0	98236	0.1567	0.0820	0.0201
0.0048	42.0	100632	0.1507	0.0817	0.0200
0.0043	43.0	103028	0.1520	0.0817	0.0202
0.0043	44.0	105424	0.1522	0.0836	0.0206
0.0037	45.0	107820	0.1559	0.0801	0.0198
0.0036	46.0	110216	0.1588	0.0797	0.0197
0.0036	47.0	112612	0.1562	0.0788	0.0196
0.003	48.0	115008	0.1674	0.0791	0.0196
0.0031	49.0	117404	0.1682	0.0787	0.0197
0.0029	50.0	119800	0.1612	0.0787	0.0197
0.0029	51.0	122196	0.1548	0.0802	0.0201
0.0025	52.0	124592	0.1630	0.0771	0.0192
0.0023	53.0	126988	0.1654	0.0792	0.0198
0.0021	54.0	129384	0.1754	0.0780	0.0193
0.002	55.0	131780	0.1687	0.0777	0.0193
0.002	56.0	134176	0.1746	0.0764	0.0191
0.0018	57.0	136572	0.1655	0.0765	0.0192
0.0019	58.0	138968	0.1688	0.0794	0.0200
0.0017	59.0	141364	0.1681	0.0753	0.0189
0.0015	60.0	143760	0.1782	0.0767	0.0191
0.0014	61.0	146156	0.1772	0.0773	0.0194
0.0014	62.0	148552	0.1839	0.0752	0.0189
0.0014	63.0	150948	0.1781	0.0765	0.0190
0.0013	64.0	153344	0.1809	0.0764	0.0192
0.0012	65.0	155740	0.1793	0.0767	0.0192
0.0011	66.0	158136	0.1766	0.0766	0.0194
0.001	67.0	160532	0.1832	0.0747	0.0186
0.001	68.0	162928	0.1845	0.0762	0.0191
0.0009	69.0	165324	0.1944	0.0747	0.0186

Framework versions

Transformers 4.50.3
Pytorch 2.6.0+cu124
Datasets 3.5.0
Tokenizers 0.21.1

📄 License

This project is licensed under the MIT 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