What is a Hidden Markov Model (HMM) in speech recognition?
A Hidden Markov Model is a probabilistic model that represents speech as a sequence of hidden states (such as phonemes) that generate observable acoustic features over time.
Why are HMMs well suited for speech recognition?
HMMs model temporal sequences and variability in speech by accounting for time-dependent transitions between speech states.
What are the hidden states in an HMM-based speech recognizer?
Hidden states typically represent subword units such as phonemes or parts of phonemes that are not directly observable.
What does the observation model in an HMM represent?
It represents the probability of observing a particular acoustic feature vector given a specific hidden state.
What does the transition model in an HMM represent?
It represents the probability of moving from one speech state to another over time.
What is a context-dependent model in speech recognition?
A context-dependent model represents speech sounds while considering their neighboring sounds, rather than modeling each sound in isolation.
Why are context-dependent models more accurate than context-independent models?
Because the pronunciation of a sound depends on surrounding sounds due to coarticulation effects in natural speech.
What is an example of a context-dependent unit?
A triphone, which models a phoneme together with its left and right neighboring phonemes.
What problem do context-dependent models help solve?
They reduce ambiguity and improve recognition accuracy by capturing phonetic variation caused by context.
What is a language model in speech recognition?
A language model estimates the probability of word sequences and helps determine which word sequence is most likely given the recognized sounds.
Why is a language model necessary in ASR systems?
Acoustic information alone is ambiguous, and the language model helps resolve this ambiguity using linguistic knowledge.
What is an n-gram model?
An n-gram model is a statistical language model that estimates the probability of a word based on the previous n−1 words.
What is a unigram model?
A unigram model assumes all words are independent and estimates word probabilities without context.
What is a bigram model?
A bigram model estimates the probability of a word based on the immediately preceding word.
What is a trigram model?
A trigram model estimates the probability of a word based on the two preceding words.
What is a limitation of n-gram language models?
They require large amounts of data and have limited ability to capture long-range dependencies.
What is a context-free grammar (CFG)?
A context-free grammar is a rule-based formalism that defines valid sentence structures using production rules.
How are context-free grammars used in speech recognition?
They constrain possible sentence structures, reducing recognition errors by enforcing grammatical rules.
What is an advantage of using CFGs in ASR systems?
They provide strong syntactic constraints and improve accuracy in limited-domain applications.
What is a limitation of context-free grammars in speech recognition?
They are difficult to scale to large, open-domain speech due to complexity and rigidity.
How do acoustic models and language models work together in ASR?
The acoustic model estimates how likely sounds match words, while the language model estimates how likely word sequences are.
Why is combining HMMs with language models effective?
HMMs handle temporal acoustic variation, while language models provide linguistic context to resolve ambiguity.
In which scenarios are grammar-based language models preferred?
In restricted domains such as voice-controlled systems with fixed command structures.
What is the overall goal of using multiple models in speech recognition?
To improve accuracy by combining acoustic, phonetic, and linguistic knowledge.
Last changed12 days ago