Historical introduction Mathematical background (e.g., pattern classification, acoustics) Feature extraction for speech recognition (and some neural processing) What sound units are typically defined Audio signal processing topics (pitch extraction, perceptual audio coding, source separation, music analysis) Now – back to pattern recognition, but include time
ASR = static pattern classification + sequence recognition ASR = static pattern classification + sequence recognition Deterministic sequence recognition: template matching Templates are typically word-based; don’t need phonetic sound units per se Still need to put together local distances into something global (per word or utterance)
Basic approach the same for deterministic, statistical: - 25 ms windows (e.g., Hamming), 10 ms steps (a frame)
- Some kind of cepstral analysis (e.g., MFCC or PLP)
- Cepstral vector at time n called xn
Words, phones most common Words, phones most common For template-based ASR, mostly words For template-based ASR, local distances based on examples (reference frames) versus input frames
Easy if local matches are all correct (never happens!) Easy if local matches are all correct (never happens!) Local matches are unreliable Need measure of goodness of fit Need to integrate into global measure Need to consider all possible sequences
Matrix for comparison between frames Matrix for comparison between frames Word template = multiple feature vectors Reference template = Input template = Need to find D( , )
Time Normalization Which references to use Defining distances/costs Endpoints for input templates
Linear Time Normalization Linear Time Normalization Nonlinear Time Normalization – Dynamic Time Warp (DTW)
Speech sounds stretch/compress differently Speech sounds stretch/compress differently Stop consonants versus vowels Need to normalize differently
Permit many more variations Permit many more variations Ideally, compare all possible time warpings Vintsyuk (1968): use dynamic programming
Bellman optimality principle (1962): optimal policy given optimal policies from sub problems Bellman optimality principle (1962): optimal policy given optimal policies from sub problems Best path through grid: if best path goes through grid point, best path includes best partial path to grid point Classic example: knapsack problem
Stuffing a sack with items, different value Stuffing a sack with items, different value Goal: maximize value in sack Key point 1: If max size is 10, and we know values of solutions for max size of 9, we can compute the final answer knowing the value of adding items. Key point 2: Point 1 sounds recursive, but can be made efficiently nonrecursive by building a table
Apply DP to ASR: Vintsyuk, Bridle, Sakoe Apply DP to ASR: Vintsyuk, Bridle, Sakoe Let d(i,j) = local distance between frame i in input and frame j in reference Let p(i,j) = set of possible predecessors to frame i in input and frame j in reference D(i,j) = d(i, j) + minp(i,j) D(p(i,j))
(1) Compute local distance d in 1st column(1st frame of input) for each reference template. Let D(0,j) = d(0,j) for each cell in each template (2) For i=1 (2nd column), j=0, compute d(i,j) add to min of all possible predecessor values of D to get local value of D; repeat for each frame in each template. (3) Repeat (2) for each column to the end of input (4) For each template, find best D in last column of input (5) Choose the word for the template with smallest D
O(Nframesref . Nframesin . Ntemplates) O(Nframesref . Nframesin . Ntemplates) Storage, though can just be O(Nframesref . Ntemplates) (store current column and previous column) Constant reduction: global constraints Constant reduction: local constraints
All examples? All examples? Prototypes? DTW-based global distances permit clustering
(1) Initialize (how many, where) (1) Initialize (how many, where) (2) Assign examples to closest center (DTW distance) (3) For each cluster, find template with minimum value for maximum distance, call it the center (4) Repeat (2) and (3) until some stopping criterion is reached (5) Use center templates as references for ASR
Normalizing for scale Cepstral weighting Perceptual weighting, e.g., JND Learning distances, e.g., with ANN, statistics
Sounds easy Sounds easy Hard in practice (noise, reverb, gain issues) Simple systems use energy, time thresholds More complex ones also use spectrum Can be tuned Not robust
Time normalization Time normalization Recognition Segmentation Can’t have templates for all utterances DP to the rescue
Vintsyuk, Bridle, Sakoe Vintsyuk, Bridle, Sakoe Vintsyuk, Bridle: one stage Ney explanation Ney, H., “The use of a one-stage dynamic programming algorithm for connected word recognition,” IEEE Trans. Acoust. Speech Signal Process. 32: 263-271, 1984
In principle: one big distortion matrix (for 20,000 words, 50 frames/word, 1000 frame input [10 seconds] would be 109 cells!) In principle: one big distortion matrix (for 20,000 words, 50 frames/word, 1000 frame input [10 seconds] would be 109 cells!) Also required, backtracking matrix (since word segmentation not known) Get best distortion Backtrack to get words Fundamental principle: find best segmentation and classification as part of the same process, not as sequential steps
In principle, backtracking matrix points back to best previous cell In principle, backtracking matrix points back to best previous cell Mostly just need backtrack to end of previous word Simplifications possible
Distortion matrix -> 2 columns Distortion matrix -> 2 columns Backtracking matrix -> 2 rows “From frame” points to end frame of previous word
“Within word” local constraints “Within word” local constraints “Between word” local constraints Grammars Transition costs
DTW combines segmentation, time norm, recognition; all segmentations considered DTW combines segmentation, time norm, recognition; all segmentations considered Same feature vectors used everywhere Could segment separately, using acoustic-phonetic features cleverly Example: FEATURE, Ron Cole (1983)
No structure from subword units No structure from subword units Average or exemplar values only Cross-word pronunciation effects not handled Limited flexibility for distance/distortion Limited mathematical basis -> Statistics!
Having examples can get interesting again when there are many of them Having examples can get interesting again when there are many of them Potentially an augmentation of stat methods Recent experiments show decent results Somewhat different properties -> combination
Statistical ASR Statistical ASR Speech synthesis Speaker recognition Speaker diarization Oral presentations on your projects
Week of April 30: no class Monday, double class Wednesday May 2 (is that what people want?) Week of April 30: no class Monday, double class Wednesday May 2 (is that what people want?) 8 oral presentations by individuals, 12 minutes each + 3 minutes for questions 2 oral presentations by pairs – 17 minutes each + 3 minutes for questions 3:10 PM to 6 PM with a 10 minute mid-session break Written report due Wednesday May 9, no late submissions (email attachment is fine)
Do'stlaringiz bilan baham: |