Use the Speech Recognition Engine's
API in 5 easy steps!

1. Open a new Speech Port representing your connection to the Speech Engine.
2. Load a grammar — multiple grammars can be loaded at once.
3. Load the audio data into a sound channel.
4. Tell our Engine to do a decode on the audio.
5. Get the recognized results!

Engine Functionality

• Streaming Audio
• Support for US, UK, Australian & New Zealand English, Mexican and South American Spanish, & Canadian French.
• Flexible API with support for MRCP v1 and 2 (NLMSL).
• Client/Server architecture distributes speech–processing load.
• Runtime defined grammars entered as simple text, BNF, raw phonetic spelling or SRGS.
• Advanced dynamic barge–in adapts to each call in real time.
• SDK includes examples of demo apps.
• Flexible error recovery through the use of confidence scores and NBest results.

How Speech Recognition Works

Speech engines use this process to figure out what a speaker said:

The Speech Engine compares audio with the loaded grammars
to produce recognized text.

Licensing

The Speech Engine is licensed per port for the duration of the interaction. Call for details.

LumenVox offers tools, support and training to ensure your speech application success.

We can get you started quickly. Please contact us to discuss your specific application needs.

LumenVox speech recognition technology proudly supports and complements the following standards. LumenVox consistently works toward collaboration standards in the speech industry to develop solutions for business.

MRCP

Media Resource Control Protocol

Speech Synthesizers. Audio Recorders. DTMF Recognizers. Speech Recognizers. Speech Verifiers. A fully functioning, media–rich application needs a lot of components to work together. Until now, all of these components were provided by a single vendor, or required extensive custom programming integration. Media Resource Control Protocol (MRCP ) changes all this. MCRP allows you to seamlessly manage diverse media resources providing a common language to speak to all of these devices.

With MRCP, vendors can compete on the basis of their strengths, rather than one vendor attempting to create an all–inclusive potentially mediocre package. Therefore, you're empowered to integrate the best vendor product, allowing you to create a speech application package tailored to your specific needs. MRCP is a protocol designed to address media processing resources control, i.e. Speech Recognition and Text–To–Speech (TTS) engines. LumenVox supports MRCP v1 and 2 (returning results in NLMSL).

SISR

Semantic Interpretation for Speech Recognition

LumenVox has implemented the W3C's Semantic Interpretation for Speech Recognition (SISR) recommendation, which is also part of the VXML 2.0 specification. SISR allows grammar authors to embed snippets of JavaScript code into their SRGS grammars to automatically transform what a speaker says into a format understandable to an application. With LumenVox's Semantic Tags, callers can say, "September thirteenth two thousand eleven," and your application will understand "2011–09–13."

SISR

Semantic Interpretation for Speech Recognition

The W3C defined a syntax called the Speech Recognition Grammar Specification (SRGS), for representing grammars for use in speech recognition so that developers can specify the words and patterns of words to be listened for by a speech recognizer. The syntax of the grammar format is presented in two forms, an Augmented BNF Form and an XML Form. The specification makes the two representations mappable to allow automatic transformations between the two forms. The LumenVox Speech Engine supports the SRGS, as defined by the W3C. More information visit: www.w3.org/TR/speech-grammar/

VXML

Voice Extensible Markup Language

Voice Extensible Markup Language (VXML) is a mark–up language designed to code speech applications with many of the same architectural components as HTML. VoiceXML platforms connect to a combination of speech recognition engines, text–to–speech synthesis, telephony interfaces and a VoiceXML Interpreter software to process the call. To interface VXML with any speech engine, the engine must understand SRGS and SISR.

LumenVox's Speech Engine is compliant with what VXML. The Lumenvox Speech Engine powers the speech recognition portion of many VXML platforms.

VXML Forum

Voice Extensible Markup Language

The VoiceXML Forum is an industry organization formed to create and promote VoiceXML. With the backing and contributions of its diverse membership, including key industry leaders, the VoiceXML Forum has successfully driven market acceptance of VoiceXML through a wide array of speech–enabled applications. LumenVox is a proud member of the Forum. For more information visit: www.voicexml.org.

Server–Side Speech Grammars

LumenVox offers even more efficient support for large speech recognition grammars through advanced server– and client–side grammar caching. Server–side speech grammars allow large grammars to be compiled once and saved so that they can be loaded again very quickly and without using up valuable processor or memory resources.

Voice Activity Detection

In our mobile society, we rarely make calls while sitting in a quiet room. Whether the call is coming from a crowded restaurant or inside of a speeding car, separating speech from background noise is a tricky task for speech recognition software.

The LumenVox Speech Engine uses a technology called Voice Activity Detection (VAD) to distinguish between actual speech and other sounds. Human speech has qualities that make it distinguishable from other sounds. VAD listens to the incoming audio for these qualities. These include:

• Energy Level (volume)
• Frequency (pitch)
• Changes in frequency
• Duration

NBest Results

Instead of returning only the top scoring result, you can instruct the Speech Engine to return several of the highest scoring, most likely answers, often called NBest results. Returning NBest results is particularly effective when callers need to spell names, street addresses, or e–mail addresses. Without NBest results, if a caller spells a name beginning with "N," but the engine returns a low confidence score, the caller would be asked to repeat the letter — and given how similar "N" is to "M," it's likely that the second answer would have a similarly low confidence score. With NBest results, the system can prompt the caller using several of the likely results, such as "Did you mean 'M,' as in 'Mary'?" When the caller responds, "No," the system goes to its next option, "Perhaps you meant 'N,' as in 'Nancy'?"

Speech Engine Sample Code

Speech Engine Comparison

We are often asked, "How accurate is the LumenVox Speech Engine?" In response, we publish our Acoustic Models for the Speech Engine which validates improvements in accuracy and shows comparative data to our clients and partners.

Speech Engine Comparison

We are often asked, "How accurate is the LumenVox Speech Engine?" In response, we publish our Acoustic Models for the Speech Engine which validates improvements in accuracy and shows comparative data to our clients and partners.

What Is Accuracy?

Accuracy is the correct recognition of in–grammar utterances, or how successful the speech recognition engine is at recognizing a word or phrase that is an expected spoken response.

How Did We Compare?

For our Accuracy Study we compared the LumenVox Speech Engine with another leading Speech Engine, "Competitor N," using third–party test data. We tested five different utterance types, or "domains," which were: Yes/No, Name & Department (ex. Company Directory), Date (Day, Month & Year), Numbers (such as two thousand, one hundred) and Digits (a string of digits).

It should be noted that the test set is a raw sampling of data from telephones and cell phones and includes a reasonable percentage of "noisy" data.

Results

Our Speech Engine fared extremely well, with comparatively high accuracy marks in almost all categories.

What Does This Mean?

After reviewing the analysis, you can be confident of the accuracy, performance and position of LumenVox speech recognition software in the speech industry.

"In order to get any Technical Support with the previous vendor, you have to go through their process, and pay their prices. And even so, they were not really giving us any solutions."
Gustavo Berdinas

*All engines intentionally not optimized for utterances, applications or grammars. In addition, in order to ensure an objective test, we made no effort to tune or make adjustments to the software.

Tuned Versus Untuned Applications

The accuracy graphs above only reflect accuracy for untuned applications. The tested applications did not have their grammars — the list of words that will be recognized by a speech engine — optimized based on actual call data.

Because those accuracy figures are for untuned applications, they do not represent the final accuracy results users can expect from deployed speech recognition applications.

The graph to the right shows the accuracy results for the speech–enabled call router LumenVox uses. The router contains approximately 70 names.

Before being tuned, the router contained only proper names. After examining call data with our Speech Tuner, LumenVox added into the grammar alternative pronunciations and nicknames of employees. As you can see, this relatively simple act of tuning caused accuracy to move from 85.54% to 96.21%.

LumenVox Sizing / Hardware Requirements

When building speech applications we're often asked, "How many ports or channels can I run on a given machine?"

In general, the number of simultaneous speech sessions that can be reliably run varies with the capacity of the machine and the type of speech recognition being performed. For instance, recognizing a single digit is computationally less "expensive" than recognizing a full street address from a complex grammar.

At LumenVox, we can provide an estimate of the number of ports or channels you'll need by designing an application that mirrors the most common speech application and running it on a variety of hardware configurations. Our sample application simulates a user calling into an IVR that performs both speech and DTMF interactions. The call progresses through a series of interactions, including digits, currencies, date/time, menu commands and mixed grammars.

With specialized testing tools, we're able to generate large numbers of calls at the same time and measure the performance of LumenVox speech software to determine numbers that we use as guidelines for developers building standard speech IVRs. The tests are conducted in Windows and Linux configurations:

Low resource machine	85 concurrent sessions
High resource machine	360 concurrent sessions

These results are intended as a guide. Your application may allow for significantly higher or lower density.

Test Environment

The test environment configuration:

Low resource machine
Intel Xeon CPU, 2.4 GHz (), total of 4 CPU cores
2.0 GB of RAM
32–bit operating system (Windows XP Professional SP2 32–bit and CentOS 5.4 Linux 32–bit)

High resource machine
2x Intel Xeon CPU, 2.33 GHz (E5410), quad-core, for a total of 8 CPU cores
8.0 GB of RAM
64–bit operating system (Windows Server 2008R2 64–bit and CentOS 5.4 Linux 64–bit)

Test configuration

• LumenVox Version 10.2
• LumenVox stack (Media Server, Speech Engine, License Server) installed on the target machine
• Calls driven by a separate machine on the same local area network as the LumenVox server
• When call finishes, the system waits 10 seconds before sending the next call

NOTE: To reach the higher levels seen in the test, some default LumenVox configurations were changed. For example, by default the Media Server is only configured to use up to 200 ports. This obviously needs to be increased to get 360+ simultaneous sessions.

Please contact LumenVox for more information if you are attempting to run higher density sessions.

Test criteria

The LumenVox test seeks to measure performance of the system in the same way a person would when calling into an IVR. Thus we did not measure hardware load directly, but instead ensured the system performed exactly as expected. Our criteria for a successful test are:

• The system must perform at greater than 95% success rate for all events
• An "event" consists of a speech recognition, DTMF interaction, or new call
• The average time to initiate a new session must also have been less than 1 second
• For a speech or DTMF interaction to be successful, LumenVox must return the correct response within 5 seconds of the end of speech/DTMF
• Interactions had a 70%/30% split of speech/DTMF

LumenVox Speech Engine

The hardware resources required by speech recognition software vary significantly depending on the exact task being performed. The main variables involve the length of audio, total number of simultaneous channels of recognition, and the type of application (specifically the size and complexity of the grammars being used).

For production telephony systems, LumenVox recommends a modern server running a 64–bit operating system with 4–8 CPU cores and 6 GB of RAM.

However, different applications and use cases may have significantly lower or higher requirements. Please contact us for help sizing your specific hardware needs.

Version 10.2.100

Released 7/22/2011

• New LumenVox Manager Service
• New LumenVox Dashboard application
• Monitors LumenVox Services, incl. option to stop/start them
• Displays centralized application logs and statistics
• Manages licensing
• LvMediaServerMonitor and LVLicenseAdministrator no longer supported after 10.2
• Improved centralized logging

	--OLD--	--NEW--
	LVStatus_TTSServer.log	tts_server_status.txt
	LVStatus_SREServer.log	asr_server_status.txt
	LVStatus_LVLicenseServer.log	license_server_status.txt
	LVStatus_MediaServer.log	media_server_status.txt
	LVApp_SRE.log	asr_server_app.txt
	SpeechServerLog.txt	asr_server_app.txt
	LVApp_SpeechPort.log	client_asr.txt
	---	client_tts.txt
	---	client_license.txt
	LVApp_LicenseServer.log	license_server_app.txt
	LvMediaServer.txt	media_server_app.txt
	LVApp_Tuner.log	tuner_app.txt
	LVApp_Tuner.log	call_indexer_app.txt
	LVApp_TTSServer.log	tts_server_app.txt
	---	dashboard_app.txt
	---	manager_app.txt
	GrammarLoadLog.txt	asr_server_grammar.txt
	LVApp_Critical.log	lumenvox_critical.txt
	LVApp_MediaServer_Restart.log	media_server_restart.txt
	LVApp_SREServer_Restart.log	asr_server_restart.txt
	LVApp_LicenseServer_Restart.log	license_server_restart.txt
	LVApp_TTSServer_Restart.log	tts_server_restart.txt
	---	manager_restart.txt




• Revised memory management — 40% of previous use when under load
• Addition of SSML pre–parser
• Faster and improved Answering Machine detection
• New Fax Tone Detection
• Answering Machine Detection support has been added to the LumenVox Tuner
• Improved acoustic models for U.S. and British English and Mexican Spanish



	en–GB	en–IN	es–MX
v10.1	80.65 (96.12)	82.77 (95.72)	84.53 (95.08)
v10.2	87.74 (97.91)	84.31 (96.08)	86.19 (94.72)



• Phonetic speller tool now has history and cut/paste options