Open-source Speech Recognition Model wav2vec2-xls-r-300m-zh-TW - Accurately Recognize Taiwanese Chinese Speech

Wav2vec2 Xls R 300m Zh TW

Developed by StevenLimcorn

This is a Chinese-Taiwan speech recognition model fine-tuned on the COMMON_VOICE - ZH-TW dataset based on the facebook/wav2vec2-xls-r-300m model

Speech Recognition

Transformers

Open Source License:Apache-2.0 #Taiwanese Mandarin Speech Recognition #High Accuracy Speech-to-Text #Common Voice Dataset

Downloads 58

Release Time : 3/2/2022

Model Overview

This model is specifically designed for Chinese-Taiwan speech recognition tasks and has been fine-tuned on the Common Voice dataset

Model Features

Chinese-Taiwan Speech Recognition

Optimized specifically for Chinese-Taiwan accent speech recognition

Based on wav2vec2-xls-r Architecture

Uses the powerful wav2vec2-xls-r-300m pre-trained model as the foundation

Fine-tuned on Common Voice Dataset

Fine-tuned on the Common Voice Chinese-Taiwan dataset

Model Capabilities

Chinese-Taiwan Speech Recognition

Speech-to-Text

Automatic Speech Recognition

Use Cases

Speech Transcription

Meeting Minutes Transcription

Convert Chinese-Taiwan accent meeting recordings into text transcripts

Word Error Rate (WER): 0.8594, Character Error Rate (CER): 0.2964

Voice Assistant

Provide voice input support for Chinese-Taiwan users

Accessibility Applications

Real-time Caption Generation

Provide real-time caption services for the hearing impaired

🚀 Automatic Speech Recognition Model

This is a fine - tuned speech recognition model based on the Transformer architecture, which can achieve high - quality speech recognition on the COMMON_VOICE - ZH - TW dataset.

🚀 Quick Start

This model is a fine - tuned version of [facebook/wav2vec2 - xls - r - 300m](https://huggingface.co/facebook/wav2vec2 - xls - r - 300m) on the COMMON_VOICE - ZH - TW dataset. It achieves the following results on the evaluation set:

Loss: 1.1786
Wer: 0.8594
Cer: 0.2964

📚 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: 7.5e - 05
train_batch_size: 8
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 32
optimizer: Adam with betas=(0.9,0.999) and epsilon = 1e - 08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 2000
num_epochs: 100.0
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
64.6189	2.51	500	63.8077	1.0	1.0
8.0561	5.03	1000	6.8014	1.0	1.0
6.0427	7.54	1500	6.0745	1.0	1.0
5.9357	10.05	2000	5.8682	1.0	1.0
5.0489	12.56	2500	4.4032	0.9990	0.7750
4.6184	15.08	3000	3.8383	0.9983	0.6768
4.365	17.59	3500	3.4633	0.9959	0.6299
4.1026	20.1	4000	3.0732	0.9902	0.5814
3.8655	22.61	4500	2.7638	0.9868	0.5465
3.6991	25.13	5000	2.4759	0.9811	0.5088
3.4894	27.64	5500	2.2937	0.9746	0.4852
3.3983	30.15	6000	2.1684	0.9733	0.4674
3.2736	32.66	6500	2.0372	0.9659	0.4458
3.1884	35.18	7000	1.9267	0.9648	0.4329
3.1248	37.69	7500	1.8408	0.9591	0.4217
3.0381	40.2	8000	1.7531	0.9503	0.4074
2.9515	42.71	8500	1.6880	0.9459	0.3967
2.8704	45.23	9000	1.6264	0.9378	0.3884
2.8128	47.74	9500	1.5621	0.9341	0.3782
2.7386	50.25	10000	1.5011	0.9243	0.3664
2.6646	52.76	10500	1.4608	0.9192	0.3575
2.6072	55.28	11000	1.4251	0.9148	0.3501
2.569	57.79	11500	1.3837	0.9060	0.3462
2.5091	60.3	12000	1.3589	0.9070	0.3392
2.4588	62.81	12500	1.3261	0.8966	0.3284
2.4083	65.33	13000	1.3052	0.8982	0.3265
2.3787	67.84	13500	1.2997	0.8908	0.3243
2.3457	70.35	14000	1.2778	0.8898	0.3187
2.3099	72.86	14500	1.2661	0.8830	0.3172
2.2559	75.38	15000	1.2475	0.8851	0.3143
2.2264	77.89	15500	1.2319	0.8739	0.3085
2.196	80.4	16000	1.2218	0.8722	0.3049
2.1613	82.91	16500	1.2093	0.8719	0.3051
2.1455	85.43	17000	1.2055	0.8624	0.3005
2.1193	87.94	17500	1.1975	0.8600	0.2982
2.0911	90.45	18000	1.1960	0.8648	0.3003
2.0884	92.96	18500	1.1871	0.8638	0.2971
2.0766	95.48	19000	1.1814	0.8617	0.2967
2.0735	97.99	19500	1.1801	0.8621	0.2969

Framework versions

Transformers 4.17.0.dev0
Pytorch 1.10.2+cu102
Datasets 1.18.2.dev0
Tokenizers 0.11.0

📄 License

This project 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