Voice BiometricsVoice Biometrics

General DescriptionGeneral Description Each individual has individual voice Each individual has individual voice

components called components called phonemesphonemes.. Each phoneme has a Each phoneme has a pitchpitch, , cadencecadence, ,

and and inflection inflection These three give each one of us a These three give each one of us a

unique voice sound.unique voice sound. The similarity in voice comes from The similarity in voice comes from

cultural and regional influences in cultural and regional influences in the form of accents.the form of accents.

General DescriptionGeneral Description According to the National Center of Voice and Speech, as one According to the National Center of Voice and Speech, as one

phonate, the vocal folds and produces a complex sound phonate, the vocal folds and produces a complex sound spectrum made up of a range of frequencies and overtones. spectrum made up of a range of frequencies and overtones. As the spectrum travels through the various-sized areas in the As the spectrum travels through the various-sized areas in the vocal track, some of the frequencies resonate more than vocal track, some of the frequencies resonate more than others.others. Larger spaces resonate at a lower frequenciesLarger spaces resonate at a lower frequencies Smaller at higher frequenciesSmaller at higher frequencies

The two largest spaces in the vocal track and, the throat, and The two largest spaces in the vocal track and, the throat, and the mouth, produce the two lowest resonant frequencies or the mouth, produce the two lowest resonant frequencies or formants. formants.

Certain inflections and pitches we learn from family Certain inflections and pitches we learn from family members.members.

Voice physiological and behavior biometric are influenced by Voice physiological and behavior biometric are influenced by our body, environment, and age.our body, environment, and age.

It is possible that our voice does not always sound the same.It is possible that our voice does not always sound the same. So is voice a good biometric?So is voice a good biometric?

FormantsFormants are the resonant frequencies of the are the resonant frequencies of the vocal tract when vowels are pronounced. While vocal tract when vowels are pronounced. While vowels are attributed to this periodic resonance, vowels are attributed to this periodic resonance, consonants are not periodic. They are produced consonants are not periodic. They are produced by restriction of air flow with the mouth, tongue, by restriction of air flow with the mouth, tongue, and jaw.and jaw.

Linguists classify each type of speech sound Linguists classify each type of speech sound (called phenomes) into different categories. In (called phenomes) into different categories. In order to identify each phenome, it is oftentimes order to identify each phenome, it is oftentimes useful to look at its spectrogram or frequency useful to look at its spectrogram or frequency response where one can find the characteristic response where one can find the characteristic formants formants

General DescriptionGeneral Description

Although all phenomes have their own Although all phenomes have their own formants, vowel sound formants are usually formants, vowel sound formants are usually the easiest to identify the easiest to identify

All formants have the trait of waxing and All formants have the trait of waxing and waning in energy in all frequencies, which is waning in energy in all frequencies, which is caused by the repeated closing and opening of caused by the repeated closing and opening of the human vocal tract. the human vocal tract. On average, this On average, this repeated closing and opening occurs at a rate repeated closing and opening occurs at a rate of 125 times per second in an adult male and of 125 times per second in an adult male and 250 times per second in an adult female.250 times per second in an adult female.

This rate gives the sensation of pitch (higher This rate gives the sensation of pitch (higher frequencies result in higher pitches). frequencies result in higher pitches).

Formant values Formant values can vary widely from person can vary widely from person to person, but the spectrogram reader learns to person, but the spectrogram reader learns to recognize patterns which are independent to recognize patterns which are independent of particular frequencies and which identify of particular frequencies and which identify the various phonemes with a high degree of the various phonemes with a high degree of reliability. reliability.

Vowel “I”Vowel “A”

Formants can be seen very clearly in a Formants can be seen very clearly in a wideband spectrogram, where they are wideband spectrogram, where they are displayed as dark bands. The darker a displayed as dark bands. The darker a formant is reproduced in the spectrogram, formant is reproduced in the spectrogram, the stronger it is (the more energy there is the stronger it is (the more energy there is there, or the more audible it is): there, or the more audible it is):

But there is a difference between oral vowels But there is a difference between oral vowels on one hand, and consonants and nasal vowels on one hand, and consonants and nasal vowels on the other. on the other.

Nasal consonants and nasal vowels can exhibit Nasal consonants and nasal vowels can exhibit additional formants, nasal formants, arising additional formants, nasal formants, arising from resonance within the nasal branch.from resonance within the nasal branch.

Consequently, nasal vowels may show one or Consequently, nasal vowels may show one or more additional formants due to nasal more additional formants due to nasal resonance, while one or more oral formants resonance, while one or more oral formants may be weakened or missing due to nasal may be weakened or missing due to nasal antiresonance. antiresonance.

Oral formants are numbered consecutively Oral formants are numbered consecutively upwards from the lowest frequency. In the upwards from the lowest frequency. In the example, fragment from the previous example, fragment from the previous wideband spectrogram shows the sequence wideband spectrogram shows the sequence [ins] from the beginning. Five formants are [ins] from the beginning. Five formants are visible in this [i], labeled F1-F5. Four are visible in this [i], labeled F1-F5. Four are visible in this [n] (F1-F4) and there is a hint visible in this [n] (F1-F4) and there is a hint of the fifth. There are four more formants of the fifth. There are four more formants between 5000Hz and 8000Hz in [i] and [n] between 5000Hz and 8000Hz in [i] and [n] but they are too weak to show up on the but they are too weak to show up on the spectrogram, and mostly they are also too spectrogram, and mostly they are also too weak to be heard. weak to be heard.

The situation is reversed in this [s], where The situation is reversed in this [s], where F4-F9 show very strongly, but there is little F4-F9 show very strongly, but there is little to be seen below F4. to be seen below F4.

Individual Differences in Individual Differences in Vowel ProductionVowel Production

There are differences in individual There are differences in individual formant frequencies attributable to: size, formant frequencies attributable to: size, age, gender, environment, and speech.age, gender, environment, and speech.

The acoustic differences that allow us to The acoustic differences that allow us to differentiate between various vowel differentiate between various vowel productions are usually explained by a productions are usually explained by a source-filter theorysource-filter theory..

The source is the sound spectrum The source is the sound spectrum created by airflow through the glottis created by airflow through the glottis which varies as vocal folds vibrate. The which varies as vocal folds vibrate. The filter is the vocal track itself- its shape is filter is the vocal track itself- its shape is controlled by the speaker.controlled by the speaker.

The three figures below (taken from The three figures below (taken from Miller) illustrate how different Miller) illustrate how different configurations of the vocal tract configurations of the vocal tract selective pass certain frequencies and selective pass certain frequencies and not others. The first shows the not others. The first shows the configuration of the vocal tract while configuration of the vocal tract while articulating the phoneme [i] as in the articulating the phoneme [i] as in the word "beet," the second the phoneme word "beet," the second the phoneme [a], as in "father," and the third [u] as [a], as in "father," and the third [u] as in "boot." Note how each configuration in "boot." Note how each configuration uniquely affects the acoustic uniquely affects the acoustic spectrum--i.e., the frequencies that are spectrum--i.e., the frequencies that are passed passed

Voice CaptureVoice Capture Voice can be captured in two ways:Voice can be captured in two ways:

Dedicated resource like a microphoneDedicated resource like a microphone Existing infrastructure like a telephoneExisting infrastructure like a telephone

Captured voice is influenced by two factors:Captured voice is influenced by two factors: Quality of the recording deviceQuality of the recording device The recording environmentThe recording environment

In wireless communication, voice travels In wireless communication, voice travels through open air and then through through open air and then through terrestrial lines, it therefore, suffers from terrestrial lines, it therefore, suffers from great interference.great interference.

Algorithms for Voice Algorithms for Voice InterpretationInterpretation

Algorithms used to capture, enroll Algorithms used to capture, enroll and match voice fall into the and match voice fall into the following categories:following categories: Fixed phase verificationFixed phase verification Fixed vocabulary verificationFixed vocabulary verification Flexible vocabulary verificationFlexible vocabulary verification Text-independent verification. Text-independent verification.

Voice VerificationVoice Verification Voice biometrics works by digitizing a Voice biometrics works by digitizing a

profile of a person's speech to produce a profile of a person's speech to produce a stored model voice print, or template. stored model voice print, or template.

Biometric technology reduces each spoken Biometric technology reduces each spoken word to segments composed of several word to segments composed of several dominant frequencies called formants. dominant frequencies called formants.

Each segment has several tones that can Each segment has several tones that can be captured in a digital format. be captured in a digital format.

The tones collectively identify the The tones collectively identify the speaker's unique voice print. speaker's unique voice print.

Voice prints are stored in databases in a Voice prints are stored in databases in a manner similar to the storing of manner similar to the storing of fingerprints or other biometric data. fingerprints or other biometric data.

Application of Voice Application of Voice TechnologyTechnology

Voice technology is applicable in a variety Voice technology is applicable in a variety of areas but for us, those used in biometric of areas but for us, those used in biometric technology include:technology include: Voice Verification Voice Verification

Internet/intranet security: Internet/intranet security: on-line banking on-line banking on-line security trading on-line security trading access to corporate databases access to corporate databases on-line information services on-line information services

PC access restriction software PC access restriction software Parental control Parental control Business software as a DSP solution at check points Business software as a DSP solution at check points

where smart cards or PIN used entrance / exit control where smart cards or PIN used entrance / exit control points points

Voice RecognitionVoice Recognition hands free devices, for example car mobile hands hands free devices, for example car mobile hands

free sets free sets electronic devices, for example telephone, PC, or electronic devices, for example telephone, PC, or

ATM cash dispenser ATM cash dispenser software applications, for example games, software applications, for example games,

educational or office software educational or office software industrial areas, warehouses, etc. industrial areas, warehouses, etc. spoken multiple choice in interactive voice response spoken multiple choice in interactive voice response

systems, for example in telephony systems, for example in telephony applications for people with disabilities applications for people with disabilities

Voice verification systems are different from voice Voice verification systems are different from voice recognition systems although the two are often confused. recognition systems although the two are often confused.

Voice recognition is used to translate the spoken word Voice recognition is used to translate the spoken word into a specific response. The goal of voice recognition into a specific response. The goal of voice recognition systems is simply to understand the spoken word, not to systems is simply to understand the spoken word, not to establish the identity of the speaker. A good familiar establish the identity of the speaker. A good familiar example of voice recognition systems is that of an example of voice recognition systems is that of an automated call center asking a user to “press the automated call center asking a user to “press the number one on his phone keypad or say the word ‘one’.” number one on his phone keypad or say the word ‘one’.” In this case, the system is not verifying the identity of In this case, the system is not verifying the identity of the person who says the word “one”; it is merely the person who says the word “one”; it is merely checking that the word “one” was said instead of another checking that the word “one” was said instead of another option. option.

Voice verification verifies the vocal characteristics Voice verification verifies the vocal characteristics against those associated with the enrolled user. against those associated with the enrolled user.

The US PORTPASS Program, deployed at remote The US PORTPASS Program, deployed at remote locations along the U.S.–Canadian border, recognizes locations along the U.S.–Canadian border, recognizes voices of enrolled local residents speaking into a voices of enrolled local residents speaking into a handset. This system enables enrollees to cross the handset. This system enables enrollees to cross the border when the port is unstaffed.border when the port is unstaffed.

How is voice recognition How is voice recognition performed?performed? Voice recognition can be divided into two classes: Voice recognition can be divided into two classes:

template matching - template matching is the simplest template matching - template matching is the simplest technique and has the highest accuracy when used properly, technique and has the highest accuracy when used properly, but it also suffers from the most limitations. but it also suffers from the most limitations.

feature analysisfeature analysis The first step is for the user to speak a word or phrase The first step is for the user to speak a word or phrase

into a microphone. into a microphone. The electrical signal from the microphone is digitized The electrical signal from the microphone is digitized

by an "analog-to-digital (A/D) converter", and is stored by an "analog-to-digital (A/D) converter", and is stored in memory.in memory.

To determine the "meaning" of this voice input, the To determine the "meaning" of this voice input, the computer attempts to match the input with a digitized computer attempts to match the input with a digitized voice sample, or template, that has a known meaning. voice sample, or template, that has a known meaning.

This technique is a close analogy to the traditional This technique is a close analogy to the traditional command inputs from a keyboard. The program command inputs from a keyboard. The program contains the input template, and attempts to match contains the input template, and attempts to match this template with the actual input using a simple this template with the actual input using a simple conditional statement. conditional statement.

The two stages of a biometric system

SoftwareSoftware Open Source Speech Software from Carnegie Open Source Speech Software from Carnegie

Mellon UniversityMellon University HephaestusHephaestus: Open Source activities at Carnegie Mellon: Open Source activities at Carnegie Mellon CMU SphinxCMU Sphinx recognition engines -- Sphinx 2, Sphinx 3, recognition engines -- Sphinx 2, Sphinx 3,

Sphinx 4, and SphinxTrain. Sphinx 4, and SphinxTrain. PocketSphinx Sphinx for embedded platforms. PocketSphinx Sphinx for embedded platforms. Festvox Project speech synthesis engines, voices and tools Festvox Project speech synthesis engines, voices and tools CMU Statistical Language Modeling Toolkit (CMU SLM) CMU Statistical Language Modeling Toolkit (CMU SLM) CMUdict -- pronunciation dictionary CMUdict -- pronunciation dictionary OpenVXI -- VoiceXML browser OpenVXI -- VoiceXML browser SALT browser - finally online! SALT browser - finally online! Audio Databases -- AN4, Microphone array, etc Audio Databases -- AN4, Microphone array, etc RavenClaw-Olympus Dialog system development toolkit. RavenClaw-Olympus Dialog system development toolkit.

We will try CMU Sphinx Group Open Source We will try CMU Sphinx Group Open Source Speech Recognition Speech Recognition http://cmusphinx.sourceforge.net/html/cmusphinx.phttp://cmusphinx.sourceforge.net/html/cmusphinx.phphp