Speaker recognition
From Wikipedia, the free encyclopedia
Speaker recognition, or voice recognition is the task of recognizing people from their voices. Such systems extract features from speech, model them and use them to recognize the person from his/her voice.
Note that strictly speaking there is a difference between voice recognition (recognizing who is speaking) and speech recognition (recognizing what is being said). Generally these two terms are frequently confused and voice recognition is used as a synonym for speech recognition instead.
Speaker recognition has a history dating back some four decades, where the output of several analog filters was averaged over time for matching. Speaker recognition uses the acoustic features of speech that have been found to differ between individuals. These acoustic patterns reflect both anatomy (e.g., size and shape of the throat and mouth) and learned behavioral patterns (e.g., voice pitch, speaking style). This incorporation of learned patterns into the voice templates (the latter called "voiceprints") has earned speaker recognition its classification as a "behavioral biometric."
Contents |
[edit] Verification versus identification
Generally, two applications of speaker recognition can be distinguished: If the speaker claims to be of a certain identity and the voice is used to verify this claim this is called speaker verification or voice authentication. On the other hand, speaker identification is the task of determining an unknown speaker's identity. In a sense speaker verification is a 1:1 match where one speaker's voice is matched to one template (and possibly a general world template) whereas speaker identification is a 1:N match where the voice is matched to N templates.
Speaker verification is usually used in applications which require secure access. The systems operate with the user's knowledge and typically require their cooperation. Speaker identification systems are more likely to operate covertly without the user's knowledge. This can for example be used to route users to the correct mailbox, identify talkers in a discussion, alert speech recognition systems of speaker changes, check if a user is already enrolled in a system, etc.
[edit] Variants of speaker recognition
Each speaker recognition system has two phases: Enrollment and test. During enrollment the speaker's voice is recorded and typically a number of features is derived to form a voice print, template, or model. In the test phase (also called verification or identification phase) the speaker's voice is matched to the templates or models.
Speaker recognition systems employ three styles of spoken input: text-dependent, text-prompted and text-independent. This relates to the spoken text used during enrollment versus test.
If the text must be the same for enrollment and test this is called text-dependent recognition. It can be divided further into two cases: The highest accuracies can be achieved if the text to be spoken is fixed. This has the advantage that the system designer can devise a text which emphasizes speaker differences. However, since the text is always the same such systems are vulnerable to Impostors. Furthermore, it is not very user friendly if all users have to remember some complex text and in addition it makes the system language dependent.
Another type of text-dependent system uses pass phrases. The user is free to pick a phrase during enrollment but must use the same phrase during test. Most speaker verification applications use this type of text-dependent input. It has the advantage that an impostor must know the pass phrase, which adds a level of security. However, such systems are still vulnerable to tape recorder attacks. To counter this, many systems allow the specification of several pass phrases. Typically, these are answers to questions. During test the system randomly asks the user one of the questions and the user must provide the correct answer. However, the number of different questions is usually rather limited and thus a patient attacker could still attempt a tape recorder attack.
In text-prompted systems the speaker is asked to speak a prompted text. In principle this could be any kind of text. This however complicates the recognition process quite a bit. On the one hand the system knows the text that is being spoken. On the other hand the system must somehow know how this randomly selected text should sound if spoken by a particular speaker. This involves a much more elaborate model than in text-dependent systems where the text is always the same. The typical implementation makes use of digits. During enrollment the speaker is asked to speak a few digit sequences which are carefully selected such that all digits occur equally often. For every digit one speaker-specific model is trained. This has the advantage that only ten such models must be trained. During test a random digit sequence is selected and the corresponding digit models are concatenated into one model per speaker. Since speakers typically do not change languages frequently, such systems can be made language-independent quite easily.
Text-independent systems are most often used for speaker identification as they require very little if any cooperation by the speaker. In this case the text during enrollment and test is different. In fact, the enrollment may happen without the user's knowledge. Some recorded piece of speech may suffice.
Since text-independent systems have no knowledge of the text being spoken only general speaker-specific properties of the speaker's voice can be used. This does limit the accuracy of the recognition. On the other hand, this approach is also completely language independent.
[edit] Technology
The various technologies used to process and store voiceprints include frequency estimation, Hidden Markov models, pattern matching algorithms, neural networks, matrix representation and decision trees. Some systems also use "anti-speaker" techniques, such as cohort models, and world models.
Ambient noise levels can impede both collection of the initial and subsequent voice samples. Noise reduction algorithms can be employed to improve accuracies. Performance degradation can result from changes in behavioral attributes of the voice and from enrollment using one telephone and verification on another telephone. Voice changes due to aging also need to be addressed by recognition systems. Some systems adapt the speaker models after each successful verification to capture such long-term changes in the voice.
Many companies market speaker recognition engines, often as part of large voice processing, control and switching systems. Capture of the biometric is seen as non-invasive. The technology needs little additional hardware by using existing microphones and voice transmission technology allowing recognition over long distances via ordinary telephones (wired or wireless).
[edit] Source
[edit] References & Bibliography
[edit] Key texts
[edit] Books
[edit] Papers
[edit] Additional material
[edit] Books
[edit] Papers
[edit] External links
 This psychology-related article is a stub. You can help Wikipedia by expanding it. |
nl:Stemherkenning
