Temporal Code Temporal Code Temporal Code (Acoustic Front-end) - PowerPoint PPT Presentation

Temporal Code

Temporal Code

Temporal Code (Acoustic Front-end)

Human Recognition

Machine Recognition RECOGNIZED UTTERANCE LANGUAGE MODELING (Back-end) � HYPOTHESIZED UTTERANCES STATISTICAL SEQUENCE RECOGNITION ACOUSTIC REPRESENTATION ACOUSTIC FRONT END SPEECH WAVEFORM

Human vs. Machine

“Top-down” Processing

Machine Training • Aurora-4 Speech Database � • Wall Street Journal (WSJO) Corpus � • Large Vocabulary Continuous Speech Recognition � • 7,138 clean speech utterances, 16kHz

Human Training • Wernicke’s Area: Speech Understanding � • Broca’s Area: Speech Production

Acoustic Model Hidden Markov Model (HMM) • Each triphone characterized by HMM consisting of 3 states, 8 Gaussian mixtures per state Transition Probability Emission Probability Density

Acoustic Model • Maximum likelihood (ML) training applied to estimate a set of context-dependent triphone acoustic models

Language Model • Standard 5k lexicon (CMU pronouncing Dictionary) • Tri-gram language model

Decoder • Single-pass Viterbi beam search-based decoder

Human Recognition � Noise-Vocoder � Tone-Vocoder

CI Recognition

Normal Hearing vs. CI • Cochlear Implant range (hatched area) compared with average normal hearing scores (filled squares)

CI vs. Machine Recognition • ASR provided most accurate simulation ever!

Machine Recognition • ASR derived by world’s best auditory scientists

Effects of Training

Effects of Training

Effects of Training

Clinical Implications • Alter Frequency Allocation � • Deactivate Interfering Electrodes � • Alter Compression Curve � • Modify Electric Pulse Width

Summary

Information Technology • 2014, HMM can now improve Hearing Science

Future Work • Design improved signal processing to mimic: � • 1) Place code of neurons � • 2) Neural Firing Rates

FAME Strategy • Frequency Amplitude Modulation Encoder

SOUND S pectral Or � U ndertone N ormalization D ecomposition