HTK应用工具包API：语音识别开发的关键技术

5星 · 超过95%的资源需积分: 10 31 浏览量更新于2024-07-28 收藏 1.05MB PDF 举报

"HTK开发API是剑桥大学Machine Intelligence Laboratory为语音识别开发设计的一套工具包，它提供了对HTK（HMM Tool Kit）的调用接口，方便开发者构建自己的语音识别程序。此API名为 ATK (Application Toolkit)，版本1.6，由Steve Young在2000/2001年基于Khe Chai Sim的学生项目基础上编写，并经过多次改进，如Linux版本的实现、N-best输出功能的恢复等。CMLLR（Conditioned Maximum Likelihood Linear Regression）支持则由Hui (KK) Ye在2005年添加并测试。 HTK开发API的1.6版是一个重要的升级，新增了对语音合成的支持和异步音频输入/输出管理。为了便于处理美国英语的合成，该版本还包含了Alan Black的Flite合成器的实现。Matt Stuttle，现就职于Toshiba Cambridge Research Lab，对Linux相关问题持续提供支持，并协助准备了这个版本的发布。此资源涉及到的关键知识点包括： 1. **语音识别**：HTK是基于隐马尔可夫模型（HMM，Hidden Markov Model）的语音识别系统，它允许开发者构建和训练HMM来识别特定的语音模式。 2. **HTK API**：ATK作为HTK的接口，允许程序员更方便地与HTK工具包交互，创建自定义的语音识别应用，无需深入理解HTK的底层实现。 3. **N-best输出**：N-best列表是语音识别系统的一种输出形式，它提供识别结果的多个可能选项，按置信度排序，有助于提高识别的准确性和鲁棒性。 4. **CMLLR**：条件最大似然线性回归是统计语音识别中的一个重要技术，它可以适应说话人的个体差异，提升识别性能。 5. **Linux支持**：除了基本的语音识别功能，ATK 1.6还实现了Linux平台的兼容，扩大了其应用范围。 6. **语音合成**：新加入的合成功能意味着HTK不仅能够识别语音，还能生成语音，这对于建立双向的语音交互系统至关重要。 7. **异步音频管理**：异步音频输入/输出管理提高了系统的响应速度，使得处理音频数据更加高效，尤其在实时或高负载的应用场景下。 8. **Flite**：这是一个轻量级的文本转语音合成器，它的集成使得HTK能够实现语音的生成，对于开发具有语音反馈的系统非常有用。 9. **项目贡献**：这个工具包的发展受益于多个项目的赞助，如欧盟资助的Talk项目和CMI赞助的SCILL项目，显示了多学科合作在科研和技术进步中的重要性。通过这些知识点，开发者可以利用HTK开发API构建出功能强大的语音识别和合成系统，适用于各种应用场景，如智能家居、自动驾驶、智能助手等。

As the above illustrates, a simple application of using ATK requires very little code to implement.

For more realistic tasks, the main extension to the above code will often be to change the grammar

prior to recognising each utterance. The simplest way to do this is to load multiple grammars and then

switch them. For example, if an application required three grammars G1, G2, and G3, they would be

instantiated as objects as in the example above, and each assigned to its own resource group

ResourceGroup *group1 = rman->NewGroup("group1");

group1->AddHMMs(hset); group1->AddDict(dict); group1->AddGram(gram1);

ResourceGroup *group2 = rman->NewGroup("group2");

Group2->AddHMMs(hset); group2->AddDict(dict); group2->AddGram(gram2);

ResourceGroup *group3 = rman->NewGroup("group3");

Group3->AddHMMs(&hset); group3->AddDict(&dict); group3->AddGram(&gram3);

Notice that each resource group shares the same HMM set and dictionary but has its own unique

grammar. Given the above, the recogniser’s grammar can be switched by sending a command to it of

the form

arec->SendMessage("usegrp(group2)");

when the recogniser receives this command, it takes no immediate action. However, next time that the

recogniser is re-primed to start a new utterance, it will load the new grammar. This example illustrates

the simplest form of grammar modification. ATK allows multiple grammar objects to be stored in a

resource group and then combined in parallel. It also allows grammars to be copied, saved/restored,

edited and created from scratch.

The above overview is intended to give a flavour of what ATK is, and how it is used. The

remainder of this manual describes ATK in some detail. It starts with a description of the core classes

on which ATK is built, packets, buffers and components. It then describes the main system

components, audio source, coder, recogniser and synthesiser, followed by resources and the resource

manager. For many applications, it will be appropriate to program these directly. However, as noted

above, for spoken dialogue systems, an asynchronous i/o module called AIO can greatly simplify

design. This is described in section 8. There is then a section describing various issues relating to

HTK such as code differences, usage conventions and resources supplied with ATK. Finally, in

section 10, a few example applications are described, the source code for which is included in the

ATK distribution.

2 Packets

2.1 Generic Packet Properties

A packet is a chunk of information passed from component to component. It is polymorphic in the

sense that its inner data container can hold a variety of types. Since a packet can have multiple readers,

a reference count is also maintained for each packet instance to enable proper memory management.

Packets can therefore be passed between asynchronous threads without the need for explicit garbage

collection.

2.1.1 Programming

A Packet actually consists of 3 parts: an outer wrapper (APacket), an inner wrapper containing data-

independent property information (APacketHeader) and an inner data-specific container (a derived

class of APacketData). The outer class APacket is actually just a pointer to the inner wrapper and its

methods are forwarded to the inner wrapper. The purpose of this arrangement is to ensure that packets

can be passed around the system and properly disposed when no longer needed. Thus, no purpose is

served by creating packets with new and referencing them indirectly by pointers. Rather packets

should be instantiated as static variables and copied as needed.

The APacket class (conceptually) provides the following data members

StartTime

Start time to which the data in this packet refers

r/w

EndTime

End time of which the data in this packet refers

r/w

Kind

Kind of this packet (Wave, Observation, Phrase, ….)

Data

Pointer to inner data packet

All packet properties XXX can be read via the methods GetXXX. Properties marked as r/w can also

be written via the methods SetXXX. For example, GetStartTime() will return the start time of the

packet and SetStartTime(t) will set the start time to t (in HTK time units ie. 100ns units). Note that

start and end times make sense when the packet holds temporal data such as a segment of waveform.

In other cases, the start and end times will be the identical.

A packet is created by first creating an inner data container of the required kind e.g. for storing a

string, the following would be needed

AStringData *s = new AStringData(“A String”);

Then the packet variable itself can be declared giving the data packet as an argument to its constructor.

APacket p1(s);

Note that as mentioned above, although the inner data container is allocated using new, instantiating

the actual packet via new should be avoided.

There are a variety of packet kinds and documentation for each follows in subsequent sections.

There is also a special packet kind called AnyPacket. This is used in the context of packet filtering to

indicate that any packet kind is allowed (see ABuffer in section 3 below).

Once created, the packet may be manipulated in the usual way. For example, the start time of p1

could be printed by

printf(“Packet start time = %f\n”,p1.GetStartTime());

Copying is very cheap since only a pointer is being copied.

When copied, the inner packet (both header and the data container) is shared. Thus, after

APacket p2 (p1);

there is just one copy of the string data with two packets referencing it. The same applies for

assignment. Thus, after

p3 = p2;

the packets p2 and p3 (and p1) share the same inner packet and data container. This inner data will be

disposed of when all of p1, p2 and p3 expire. For debugging purposes, the simplest way to view a

packet is via the packet's Show() method.

2.2 Empty Packets

An empty packet contains no data at all. It is used occasionally for signalling purposes. The only

operations provided for empty packets are the generic operations described above.

2.2.1 Programming

An empty packet is created by default when the packet declaration has no initialiser. For example, in

APacket e1; APacket e2(e1)

Then e1 and e2 refer to a single empty packet. The packet’s start and end time can be set as follows

e1.SetStartTime(10000); e1.SetEndTime(25000);

Note that internally, ATK uses HTK’s 100ns time unit, hence 10000 in the above is equivalent to 1ms.

Executing the e1.Show() method would result in output of the form

Empty Packet[007B1500] refs=2; time=(1.00-2.50) ms

Data: <>

The reference in square brackets is the packet’s address; the refs value shows that the packet has two

references and that it spans the period 1.00 to 2.5 ms. This first line printed by Show() is common to

all packet types. The remainder of Show’s output depends on the particular data stored, in this case

none.

2.3 String Packets

A string packet contains a single string stored in a data container of type AStringData . It is used for

general message passing and reporting.

2.3.1 Programming

The simplest and most common way to set up a string packet is to give the required string as an

initialiser. When a packet is received, the data it contains can be accessed using the GetData method.

Since this method must be coerced to the expected type, it is usually a good idea to check the packet

kind first

APacket pkt = buffer.GetPacket(); // see Buffers section below

剩余69页未读，继续阅读

足下尺

粉丝: 1
资源: 3

HTK应用工具包API：语音识别开发的关键技术

HTK.rar_HTK_HTK UBM_htk vc_htk window_函数关系

ATK.zip_ATK_htk application

Python库 | htk-1.1.0-py2.py3-none-any.whl

语音识别代码python案例htk

matlab HTK

請給我HTK的相關文檔推薦

基于HTK（HMM）或者Kaldi框架完成一个语音识别、说话人识别或者语音合成的训练及测试

语调识别有什么可用的库？

Ç:\Users\Administrator>nc -zv 10.71.1,3 8800 nc' 不是内部或外部命令,也不是可运行的程序 或批处理文件 怎么解决？

GMM-HMM语音识别源码

最新资源

Ç:\Users\Administrator>nc -zv 10.71.1,3 8800 nc' 不是内部或外部命令,也不是可运行的程序或批处理文件怎么解决？