Speech Recognition Fundamentals

Speech Signal: Time Domain

• Speech is a sequence of different sound types
• Vowels are periodic
• Fricatives are aperiodic
• Examples include "has" and "watch"

Utterance Types

• Glides have smooth transitions, like "watch"
• Stops have transient bursts, like "dime"

Speech Signal: Frequency Domain

• Displays the speech signal as a function of frequency
• Illustrated in a graph with frequency on the x-axis and log power on the y-axis
• Shows the power spectrum of different components in the frequency domain

Automatic Speech Recognition (ASR)

• Converts speech signals into words
• Output can be used as input for natural language processing
• Recognizes speech from a speaker, converting it to words a computer can understand

Speech Recognition Process

• Input: Speech signal from a human
• Output: Text representation of the speech
• Steps include recognition, synthesis, generation and understanding of text

Speech Recognition: Main Diagram

• Signal (speech waveform) is converted to digital form
• Speech pattern is compared to models to determine units needed in the output
• The most optimal response is found using established constraints

Speech Recognition Difficulties

• Word boundary hypothesis: continuity, variability, and disfluencies in speakers
• Speaking rate variability in a number of situations
• Large vocabularies in all language and varieties
• Variability in ambient acoustics and microphone characteristics affects the ability to recognize speech in different environments
• Background noise

Speech Production/Perception

• The process of converting thoughts/ideas to a speech signal
• Diagram shows different phases involved.
• Speech production: from thoughts to acoustic signal
• Speech recognition: converting acoustic signals to understandable text and meaning
• Machine counterparts represent the systems involved, for example, printed text to the neuro-muscular movement.

Multilayer Structure of Speech Production/Recognition

• Pragmatic layer: Contextual information affecting the message
• Semantic layer: The literal meaning of the message
• Syntactic layer: The word order/syntax of the message
• Prosodic/phonetic layer: The melody and accents in the message
• Acoustic layer: The physical sound/waveform characteristics of the speech

ASR System Capabilities

• Speaking modes: range from isolated words to continuous speech
• Speaking styles: vary from read speech to spontaneous speech
• Enrollment: can be speaker-dependent or speaker-independent
• Vocabulary: varies from small to large
• SNR (signal-to-noise ratio): can range from high to low
• Transducer: from noise-cancelling microphones to cell phones

Information Extraction from Speech

• Speech signal is used to determine speaker identity
• Goal is to automatically extract information contained in a speech signal
• Speech recognition converts the speech signal to words
• Speaker recognition identifies the speaker based on their speech characteristics.

Speaker Identification

• Determines speaker identity from a known set of voices
• Closed set: all voices are known; open set: not all voices are known
• This is different from speaker verification, which determines if a claimed identity is valid.

Speaker Verification

• Synonyms: authentication, detection
• User claims an identity
• System task: to accept or reject the claimed identity
• Closed set scenario: all possible speakers are known to the system
• Impostor: All voices except the true speaker identity being matched to the claimed identity

Speaker Verification System Phases

• Enrollment phase: speech data from each speaker is collected and processed to create models
• Verification phase: a new speech sample is compared to the collected models to determine speaker identity

Verification Performance

• Many factors need to be considered, such as
• Speech quality: Channel/microphone characteristics, noise levels, and the differences in speech between enrolment and verification sessions
• Speech modality: text-based or free form speech
• Speech duration: Number of sessions of verification compared to enrollment sessions
• Speaker population size

Explore the essentials of speech signals, including time and frequency domains, along with different utterance types. Understand the process of Automatic Speech Recognition (ASR) that converts spoken language into text. This quiz covers foundational concepts crucial for anyone studying speech technology.