Gaussian Mixture Models for Acoustic Modeling
What you'll learn: How GMM-HMM systems model the probability distributions of acoustic features to recognize phonemes in traditional speech recognition.
The Core Idea
In traditional ASR, you need to answer: "Given this audio snippet, which phoneme am I hearing?" Gaussian Mixture Models (GMMs) solve this by modeling how acoustic features (like MFCCs) are distributed for each phoneme.
Think of it like this: imagine plotting all the MFCC vectors you've ever seen for the phoneme "ah" in 13-dimensional space. They don't form a single tight cluster—different speakers, pitches, and contexts create multiple overlapping clouds. A GMM represents these clouds as a mixture of several Gaussian (bell-curve) distributions, each capturing one mode of variation.
How GMMs Work with HMMs
In the GMM-HMM system you learned about previously:
- HMMs handle the temporal structure (which phoneme follows which)
- GMMs handle the acoustic likelihood (how well the observed features match a phoneme)
For each HMM state (typically representing part of a phoneme), a GMM computes the probability that the observed MFCC features came from that state. The GMM does this by:
- Maintaining multiple Gaussian components (typically 8-64 per state)
- Each Gaussian has a mean vector, covariance matrix, and mixture weight
- Combining these weighted Gaussians to model complex, multi-modal feature distributions
This flexibility lets GMMs capture variability in pronunciation, accent, gender, and recording conditions—all critical for robust phoneme recognition.
Key Takeaway: GMMs in ASR model the complex probability distributions of acoustic features for each phoneme state, enabling HMMs to compute likelihoods and recognize speech by capturing natural acoustic variability through mixtures of Gaussian components.