All Projects → nobody132 → Masr

nobody132 / Masr

Licence: other
中文语音识别; Mandarin Automatic Speech Recognition;

Programming Languages

python
139335 projects - #7 most used programming language

Projects that are alternatives of or similar to Masr

Syn Speech
Syn.Speech is a flexible speaker independent continuous speech recognition engine for Mono and .NET framework
Stars: ✭ 57 (-95.43%)
Mutual labels:  speech-recognition
Asr benchmark
Program to benchmark various speech recognition APIs
Stars: ✭ 71 (-94.3%)
Mutual labels:  speech-recognition
Sytody
a Flutter "speech to todo" app example
Stars: ✭ 79 (-93.66%)
Mutual labels:  speech-recognition
Angle
⦠ Angle: new speakable syntax for python 💡
Stars: ✭ 61 (-95.1%)
Mutual labels:  speech-recognition
Openasr
A pytorch based end2end speech recognition system.
Stars: ✭ 69 (-94.46%)
Mutual labels:  speech-recognition
Nativescript Speech Recognition
💬 Speech to text, using the awesome engines readily available on the device.
Stars: ✭ 72 (-94.22%)
Mutual labels:  speech-recognition
Biglittlenet
Official repository for Big-Little Net
Stars: ✭ 57 (-95.43%)
Mutual labels:  speech-recognition
Laibot Client
开源人工智能,基于开源软硬件构建语音对话机器人、智能音箱……人机对话、自然交互,来宝拥有无限可能。特别说明,来宝运行于Python 3!
Stars: ✭ 81 (-93.5%)
Mutual labels:  speech-recognition
Patter
speech-to-text in pytorch
Stars: ✭ 71 (-94.3%)
Mutual labels:  speech-recognition
Wav2letter
Speech Recognition model based off of FAIR research paper built using Pytorch.
Stars: ✭ 78 (-93.74%)
Mutual labels:  speech-recognition
Dragonfire
the open-source virtual assistant for Ubuntu based Linux distributions
Stars: ✭ 1,120 (-10.11%)
Mutual labels:  speech-recognition
Speech ai
Simple speech linguistic AI with Python
Stars: ✭ 66 (-94.7%)
Mutual labels:  speech-recognition
Unityandroidspeechrecognition
This repository is a Unity plugin for Android Speech Recognition (based on Java implementation)
Stars: ✭ 73 (-94.14%)
Mutual labels:  speech-recognition
Audio Pretrained Model
A collection of Audio and Speech pre-trained models.
Stars: ✭ 61 (-95.1%)
Mutual labels:  speech-recognition
Deepspeech
A PaddlePaddle implementation of ASR.
Stars: ✭ 1,219 (-2.17%)
Mutual labels:  speech-recognition
Dolphinattack
Inaudible Voice Commands
Stars: ✭ 57 (-95.43%)
Mutual labels:  speech-recognition
Android Speech Recognition
Continuous speech recognition library for Android with options to use GoogleVoiceIme dialog and offline mode.
Stars: ✭ 72 (-94.22%)
Mutual labels:  speech-recognition
B.e.n.j.i.
B.E.N.J.I.- The Impossible Missions Force's digital assistant
Stars: ✭ 83 (-93.34%)
Mutual labels:  speech-recognition
Deepspeech Websocket Server
Server & client for DeepSpeech using WebSockets for real-time speech recognition in separate environments
Stars: ✭ 79 (-93.66%)
Mutual labels:  speech-recognition
Pyspeechrev
This python code performs an efficient speech reverberation starting from a dataset of close-talking speech signals and a collection of acoustic impulse responses.
Stars: ✭ 74 (-94.06%)
Mutual labels:  speech-recognition

MASR 中文语音识别

MASR是一个基于端到端的深度神经网络中文普通话语音识别项目。

原理

MASR使用的是门控卷积神经网络(Gated Convolutional Network),网络结构类似于Facebook在2016年提出的Wav2letter。但是使用的激活函数不是ReLU或者是HardTanh,而是GLU(门控线性单元)。因此称作门控卷积网络。根据我的实验,使用GLU的收敛速度比HardTanh要快。如果你想要研究卷积网络用于语音识别的效果,这个项目可以作为一个参考。

以下用字错误率CER来衡量模型的表现,CER = 编辑距离 / 句子长度,越低越好

大致可以理解为 1 - CER 就是识别准确率。

模型使用AISHELL-1数据集训练,共150小时的录音,覆盖了4000多个汉字。工业界使用的语音识别系统通常使用至少10倍于本项目的录音数据来训练,同时使用特定场景的语料来训练语言模型,所以,不要期待本项目可以和工业界的识别效果媲美。这对于Github上任何个人项目来说都不现实,除非有更先进的技术诞生。

什么叫特定场景的语料训练的语言模型?比如你使用游戏中的语音识别,它更倾向于将你的话识别成你在玩游戏时可能说的话,比如「貂蝉被蓝打死了」。而在其他场景下,「貂蝉被蓝打死了」根本就不是一句通顺的话。不信你和一个只读过三国演义没玩过王者荣耀的人说「貂蝉被蓝打死了」,你确定ta不会反问你:「啥?貂蝉被谁打死了?lan是谁?」

在单卡GTX 1080Ti上,模型每迭代一个epoch大约需要20分钟。(实验室的CUDA版本较低,不排除更新CUDA版本后会快一些的可能。)

上图为验证集的CER随epoch的训练曲线。可以看到,目前验证集CER已经下降到11%。

图中没有显示测试集的表现。测试集的CER稍高一些,在14%。

通过外接语言模型可以将测试集的CER降低到8%。

项目目前提供的预训练模型训练了大约是100个epoch时候的,已经接近最好了。

Note that the project description data, including the texts, logos, images, and/or trademarks, for each open source project belongs to its rightful owner. If you wish to add or remove any projects, please contact us at [email protected].