Speech Recognition Technology: 2018's Complete Guide

Posted by Anna Yeatts on September 7, 2017

Speech recognition technology – a machine’s ability to identify spoken words and translate them into a machine-readable format – is here to stay.

How does speech recognition technology work? Most SRT systems use complex sets of algorithms created through acoustic and language modeling. The software maps the relationship between the sounds we make and the words we intend to say. The better those algorithms, the more accurate a piece of speech recognition technology is.

Where is speech recognition technology used?

In today’s fast-paced world, speech recognition software is almost everywhere. From the way we order household supplies, refill prescriptions over the phone, interact with digital assistants, dictate a text while we rush from one task to the next, speech recognition has become commonplace. Yes, we’re disgruntled when Siri pulls up the wrong results to our query, or we huff when Facebook Messenger doesn’t quite get what we’re trying to say. But each of these moments is a different piece of speech recognition technology cranking up its algorithms and attempting to understand what we’re trying to say.

The primary uses for speech recognition are:

medical dictation/transcription
call routing
speech-to-text processing
voice dialing
voice search
assistance for users with disabilities

Voice recognition is used in security devices and is a different type of technology.

What are the variables that determine how accurate any given speech recognition technology will be?

The lexicon:

Each SRT will have a pre-programmed vocabulary with the corresponding phonetic representations.

Variations on the lexicon based on the SRTs purpose:

Letter combinations and whole words sound differently when pronounced by different people. This is how accents and dialects throw wrenches into the works. All these variations have to be stored in the speech recognition lexicon so it knows that all those different pronunciations mean the same thing. And this needs to happen for every language in the SRT. But the starting vocabulary in the lexicon should be chosen with the end user in mind. A medical application will need medical terminology but not cooking phrases.

A medical application will need medical terminology but not cooking phrases. An SRT that will be used by Native English speakers doesn’t need to be coded for German words pronounced with an Italian accent. But it does need to be coded for English words pronounced with both German and Italian accents.

Statistics

Speech is delivered to the SRT capture device or microphone via sound waves. And waves are just that. Waves. Words aren’t delivered in pre-wrapped bundles with clear beginnings and endings. As a result, the SRT uses the surrounding words to translate each sound wave into phonetic bites. Basically, speech recognition technology involves a lot of context clues.

When more real-world data is added to the SRT, the algorithm becomes more likely to choose the correct word.

The most used algorithms are neural networks, Hidden Markov models (HMM), dynamic time warping (DTW), deep learning, and end-to-end automatic speech recognition.

The different types of speech recognition technology

Speaker dependent applications

Trained to work for your voice and only your voice
To work with a different speaker, would have to be retrained or would suffer greatly in accuracy
Single speaker models
Dictation
Smartphones/iPhones/Android
Stored locally on your device

Speaker independent applications

Trained on real-world data
Can recognize thousands of different speakers
Can recognize different accents, variations, linguistic patterns, etc.
No voice training necessary
Call management and helplines
Voice portals
Voice transcription from a recording
Stored on a server or cloud

The major speech recognition technology applications in 2018

Amazon Alexa
Baidu
Braina
Dragon Medical Practice Edition
Dragon NaturallySpeaking Professional
Dragon NaturallySpeaking Premium
Dragon NaturallySpeaking Home
e-Speaking
Fusion SpeechEMR®
Google Now
Hound
LumenVox
Microsoft Cortana
One Voice Data
Siri
SmartAction Speech IVR System
Tatzi
ViaTalk
Voice Finger

Included for each SRT:

Features
Difficulty to setup & learn
Accuracy rates
Technical support
Customer feedback & reviews

Speech recognition technology features:

Amazon Alexa –

Speaker independent
Powered by Amazon Lex
Natural language understanding (NLU) technology combined with automatic speech recognition (ASR)
Estimated to be a shade less accurate but won’t release data.
Adapts to your voice.
Operates from as far as the next room.
Needs a wake word
Uses multi-turn conversations, meaning Alexa questions the user to determine what further information is needed to determine intent
Cloud-based

Compare Google Assistant, Siri, Microsoft Cortana, and Amazon Alexa:

Baidu –

The “Google of China”
Speaker independent
Responds to voice commands
Understands both English & Mandarin
Uses Deep Speech 2
7x faster than their previous system
In several cases, Baidu was just as fast as humans and claims to transcribe three times faster than humans can type.
Chinese-language voice assistant search services for Chinese speakers visiting Japan was launched in 2008 by partnering with Willcom Inc.

Braina –

Speaker independent
Digital assistant for Microsoft Windows built around language understanding, AI, and learns from conversations
Uses natural language interface to represent meanings of text
Dictation for speech to text supports 40 languages
Pro version can transcribe pre-recorded audio
Plays songs and videos
Android app for Braina converts your phone into a wireless microphone to command your PC via wifi
Text to speech capability to read aloud
Search capabilities, calculations, dictionary, & thesaurus
Create and customize keyboard macros to automate repetitive tasks
Can convert multiple speakers to text
Can be customized to recognize certain words, create answers and templates.
Understands “most of the medical, legal and scientific terms .”

Dragon Medical Practice Edition –

Speaker independent
By Nuance Communication
Claims 99% accuracy in marketing but ranked at 80.1-87% in KLAS Accuracy testing; Claims to be 20% more accurate than previous versions and over 60% more accurate than non-medical Dragon
Updated in real time
HIPAA compliant
Comprised by Petya malware in 2017
Hidden mode allows for dictation anywhere in the EHR
Template library of macros
Use iPhone or Android devices for speech input
Vocabularies for 90+ specialties included
Hosted on Microsoft Azure
Offers both front-end and back-end solutions
Uses deep learning technology and deep neural net models

Dragon NaturallySpeaking Home –

Speaker dependent
Lower accuracy than most Dragon products
Also by Nuance Communications
Does not include voice transcription tools
Recommended for home office use
Compatible with Windows
Includes DragonPad, similar to WordPad, a word processor that can be voice controlled
Also includes free downloadable software like Open Office to dictate documents compatible with Microsoft Word
Works with popular email servers like Gmail and Hotmail and social media including Facebook and Twitter for simple commands and dictation
Can use built in laptop microphone or smartphone as microphone
Recommends using Nuance-certified Bluetooth headset for best performance
Options to automate punctuation and formatting
Comes with a recorder but users report that sections of text were skipped or not transcribed when using the included recorder
Overabundance of features may be too complex for the average user, but marketed for “speech recognition newbies”
Available for Mac as Professional Individual for Mac

Dragon NaturallySpeaking Premium –

Speaker dependent
Recommended for users spending moderate amounts of time on a computer such as teachers, students, bloggers.
Custom commands
By Nuance Communications
Punctuation and formatting will not be automatically added
Speech patterns need to be altered for Dragon Premium to properly transcribe
Does allow for verbal composition commands (ie. “insert common after … “)
Vocabulary not as advanced as other versions
Allows for words to be spelled out
Recorded dictation can be transcribed though it works better with single speaker recordings. Multiple speaker dictations are beyond the range of this software.
Dictation capture via computer microphone or microphone headset
Software preferences can be customized
Can insert frequently used text and/or graphics

Dragon NaturallySpeaking Professional –

Speaker dependent
Supports 6 languages
Developed by Nuance Communications
Well suited for users with disabilities
Records multiple user voices for transcription
Customizable commands (ie. open, close, operate programs like Excel or Word) including advanced commands like sending emails to specific end users and including email body dictation.
Search capabilities for Internet Explorer, Google Chrome, Firefox using voice commands to move cursor and insert phrases into search engines
Accent support
Less accurate at recording meetings when multiple voices speaking and background noise (as would be expected)
Vocabulary considered advanced and more words can be added
Autocorrects misspelled words
Can add punctuation and format text
Integrates with other systems like Apple Pages, Apple Keynote, Apple Numbers, Microsoft Outlook 2016, and Scrivener.
Uses deep learning
Word interpretation for this Dragon model
Enhanced capture using Mac/Apple devices and microphones
Share dictation and sync capture between Apple devices
Reviewers report slow load times especially in Word
Also reported inability to correct word recognition by teaching vocabulary and had to use backspace command to delete out spelling
Takes up large amount of memory space

e-Speaking –

Speaker dependent
No extended voice training features decreases overall accuracy
Good basic features and easy to use
Icons make learning basic commands intuitive
Navigate the internet
Commands to open/close pages
Customizable commands for simple functions
Extremely affordable
Doesn’t adapt well to learning individual voices
Runs in XP, Windows, and Vista
Small file size for minimum memory requirements
Based on SAPI (Microsoft Speech) and .NET framework
Only supports expression of data as XML
Attempts to interpret not only strong semantic intent but weaker semantic intent using Jscript statements to increase accuracy
26 different dictation commands
Inexpensive at $14 for downloadable shareware

Fusion SpeechEMR® –

Speaker independent
Uses Nuance’s SpeechMagic for front-end speech recognition
iDocview™ allows for multiple signatures
Fusion Expert allows for real time voice recognition and option for self-editing
Utilizes voice recognition
Automatic spell check
Easy to use templates and routines
Centrally managed so easy to add users, modify vocabulary, adjust system settings and formatting options
Back-end and front-end language and user profiles are synced
System is optimized with health care settings in mind
Secure encryption and password protection

Google Now –

Speaker independent
Improved accuracy in loud places
Word error rate has fallen from 23% to 4.9% in the last four years
Struggles with determining the meaning of words against the massive amount of available data
Doesn’t have a name or personality compared to others – Amit Singhal, Google’s Sr. VP in charge of related products says. “I’m not saying personality shouldn’t come, but the science to get that right doesn’t fully exist.”
Voice algorithms are becoming more adept at understanding accents as more users participate
Size is its advantage
Amazon is not included within its app index
Seven languages
Available for Windows, Mac, Linux, and Chrome OS desktops in Google Chrome, and in Search apps for Android, iOS.
Attempts to predict user behavior based on search habits
Uses Google’s Knowledge Graph project to analyze meanings and connections
Sources from any website deemed appropriate and relevant

Hound –

Speaker independent
Digital assistant app launched by SoundHound
Answers verbal questions
The app’s initial strong points include navigation, local search, weather, stocks, hotels, time zones, geography, news, photo and video search, mortgage calculation, currency conversion and flight status.
Identifies songs including hummed ones
Combines speech recognition and language understanding

LumenVox –

Speaker independent
Uses natural language understanding (NLU) through Statistical Language Models (SLM)
Used for Interactive Voice Response Systems (IVR) like call systems and auto-responders
Available in 32 and 64-bit Linux and Windows
Supports nine languages
Encourages developers to use its SRT for their own applications

Microsoft Cortana –

Speaker independent
Phone assistant built into Windows 10
Messages
Searches
Sets calendar events
In 2016, Microsoft declared parity with humans. Ie. Word error rate of 5.9%
Dictate, new add-in works with Outlook, Word, and PowerPoint for Windows
Uses Bing Speech API and Microsoft Translator
Criticism is that Cortana struggles with intent and often returns Bing search results instead of answering questions
Enables transcribing voice in more than 20 languages
Supports real time translation of up to 60 languages
Spoken commands for new lines, delete, punctuation, and more formatting
Can integrate with third party apps
Indexes and stores user information leading to privacy concerns though this option can be disabled
Disabled for users aged 13 and under
Semantic database is Satori similar to Google’s Freebase or Knowledge Graph

One Voice Data –

Speaker independent
A strategic partner of M*Modal and awarded Best in KLAS in 2017
Broadest dialectic capability of any speech recognition software on the market
Understands accents regardless of geographic, ethnic or other speech characteristics
Combines speech recognition with natural language processing to integrate word recognition with meaning, intent, and context (Speech Understanding)
Builds on collective experience of all users and individual user for increased accuracy
Paired with actual physician dictations to create collective, predictive models. Each dictation is compared to the model and checked for validity.
Complex language modeling
Accommodates all medical specialties
Encrypted and password protected
HIPAA compliant
Data exported and ready for medical coding or formatting via One Voice Data’s coding automation tool
EMR compatible
Also available as part of end-to-end automated workflow

Siri –

Speaker independent
Developed by Apple but originally a spinoff from an artificial intelligence program, Cognitive Assistant that Learns and Organizes (CALO) funded by the Defense Advanced Research Projects (DARP).
Natural language user (NLU) interface
Answers questions, make recommendations and performs tasks by delegating requests to Internet services
Most used personal assistant in the US
Adapts to individual users
Problem is nonsensical answers
Siri’s biggest hurdle is understanding spoken requests
Apple hired a team of deep learning developers in 2014 and the problem is being worked on.
Speech recognition engine powered Nuance Communications

SmartAction Speech IVR System –

Speaker independent
Fully automated voice recognition software for call centers
Customer service powered by A.I.
Intended for the middle ground between simple tasks easily accomplished by normal automation and more complex tasks that need human interaction
Specializes in complex, repetitive tasks
Cloud-based
Automatic form fill-in
Call analysis
Uses concatenated speech
Continuous speech
Customizable macros
Incorporation of specialty vocabularies
Recognizes repeat customers, remembers previous conversations, and learns from multiple interactions
Combines voice, text, and chatbot features

Tatzi –

Speaker dependent
Basic commands and features without advanced tools
Simple with adequate accuracy for speech recognition software
No additional voice training
No microphone or headset
No voice transcription
Customizable commands that are specific
Quick response times
Search function for the internet
Microphone can be turned on and off with voice command
Most inaccuracies occurred with punctuation (ie. inserting semicolons or commas) but that’s a common enough error in VRT
Text can be corrected
Cursor controlled via voice command
Switch/minimize windows and open/close programs
For Windows with emphasis on video and PC games played by voice
Created by Voice Tech Group

ViaTalk

Speaker dependent
Mobility features make it a good choice for frequent travelers
Doesn’t adapt to your voice
Lower accuracy rating than other systems
Multilingual flexibility is major asset
Can translate between languages
No advanced voice recognition software
Downloadable to smartphone or tablet
Small font size can be troublesome for some
Infrequent updates
Can purchase text scanning pen to upload text from page into a Word file. This text can be automatically translated

Voice Finger

Speaker dependent
For Windows Vista, Windows 7 & 8
Enables voice activated control of mouse cursor and keyboard
Uses a 44×44 grid to place the cursor
Improves on Windows Speech Recognition by simplifying standard voice commands
Can also add custom commands
Interface is simple and intuitive to use
Dictation primarily handled by Windows 7 built-in speech recognition
1 supported language
Improves on default Windows voice recognition settings
Main focus is for users with disabilities and injuries