Neural Sync Active
Learning Objectives
Registry Synced
Learning Objectives
427 words
2 min read
Reading compass
Now · 1. ASR Architecture
Learning Objectives
- Understand ASR system architecture
- Apply decoding and search algorithms
- Evaluate ASR performance
- Weeks 4-5: HMMs and GMMs
- Understanding of probability
1. ASR Architecture
Components:
- Front-end: Feature extraction (MFCCs)
- Acoustic Model: GMM-HMM (P(X|W) - probability of features given words)
- Pronunciation Dictionary: Word -> phoneme sequence mapping
- Language Model: Word sequence probability (P(W))
- Decoder: Combine acoustic + language models to find most likely word sequence
2. Decoding
Viterbi Beam Search: Dynamic programming search through state space. Maintains beam of active hypotheses at each time step. Prunes low-probability paths to maintain tractability.
Token Passing: Each hypothesis carries accumulated score, history (word sequence), and current HMM state.
Language Model Integration: Acoustic score + word penalty + language model score (log probabilities).
3. Performance Metrics
Word Error Rate (WER):
S = substitutions, D = deletions, I = insertions, N = reference words
Real-Time Factor (RTF): Processing time / audio duration. RTF < 1 = faster than real-time.
Q1: What are the 5 components of an ASR system?Feature extraction (MFCCs), Acoustic model (GMM-HMM/DNN), Pronunciation dictionary (word->phonemes), Language model (n-gram/neural), Decoder (Viterbi beam search). Q2: How is Word Error Rate calculated?WER = (Substitutions + Deletions + Insertions) / Reference Word Count. Minimum between reference and hypothesis using dynamic programming (Levenshtein distance). 0% = perfect, ~5% = human parity, >30% = poor. Q3: What is beam search in ASR decoding?Pruning technique: at each time step, only keep top-B hypotheses (by cumulative score). Reduces search space from exponential to O(B*T). Trade-off: narrow beam = faster but may miss correct path. Q4: What is the role of the pronunciation dictionary?Maps words to phoneme sequences (e.g., "speech" -> /s p iy ch/). Handles multiple pronunciations ("either" -> /iy dh er/ or /ay dh er/). Typically 100k+ words for English. Q5: What is a language model weight (LMW) in ASR?Scale factor applied to LM score during decoding. Optimal value (~10-20) balances acoustic and language model contributions. Higher weight = more constrained (better for narrow domain). Lower = more open (better for general). Q6: How to reduce WER?More training data, better features (MFCC + delta + speaker normalization), more context (triphones), better acoustic models (DNN), stronger language models (neural LM), speaker adaptation (fMLLR), data augmentation (noise, speed perturbation). Join Discord PreviousGaussian Mixture Models & Acoustic ModelingNextDeep Learning for ASR