Speech recognition P(words|signal)= P(signal|words) P(words) / P(signal) P(signal|words): Acoustic model P(words): Language model Idea: Maximize P(signal|words) P(words) Today: Acoustic model
Variability Variation Speaker Pronunciation Environmental Context Static acoustic model will not work in real applications. Dynamically adapt P(signal|words) while using the system.
Measuring errors (1) 500 sentences of 6 – 10 words each from 5 to 10 different speakers. 10% relative error reduction Training set / Development set First decide optimal parameter settings.
Measuring errors (2) Word recognition errors: Substitution Deletion Insertion Correct: Did mob mission area of the Copeland ever go to m4 in nineteen eighty one? Recognized: Did mob mission area ** the copy land ever go to m4 in nineteen east one?
Measuring errors (3) Correct: The effect is clear Recognised:Effect is not clear Error Rate One by one: 75% Subs + Dels + Ins #words in correct sentence Word error rate=100% x Word error rate
Units of speech (1) Modeling is language dependent.fixme Modeling unit Accurate Trainable Generalizable
Units of speech (2) Whole-word models Only suitable for small vocabulary recognition Phone models Suitable for large vocabulary recognition Problem: over-generalize less accurate Syllable models
Context dependency (1) Recognition accuricy can be improved by using context-dependent parameters. Important in fast / spontanious speech. Example: the phoneme /ee/
Context dependency (2) Triphone model: phonetic model that takes into consideration both the left and the right neightbouring phones. If two phones have the same identity, but different left or right contexts, there are considered different triphones. Interword context-dependent phones. Place in the word: Beginning Middle End
Context dependency (3) Stress Longer duration Higher pitch More intensity Word-level stress Import – Import Italy – Italian Sentence-level stress I did have dinner.
Context dependency (4) Vary much triphones. 50 3 = 125.000 Many phonemes have the same effects /b/ & /p/ labial (pronounces by using lips) /r/ & /w/ liquids Clustered acoustic-phonetic units Is the left-context phone a fricative? Is the right-context phone a front vowel?
Acoustic model After feature extraction, we have a sequence of feature vectors, such as the MFCC vector, as input data. Feature stream Phonemes / units Words Segmentation and labeling Lexical access problem
Acoustic model Signal Phonemes Problem: phonemes can be pronounced differently Speaker differences Speaker rate Microphone
Acoustic model Phonemes Words The three major ways to do this: Vector Quantization Hidden Markov Models Neural Networks
Acoustic model Problem: Multiple pronunciations: owt aa ey tow t ax m aa ey tow 0,5 0,8 m Dialect variation Coarticulation 0,5 0,2