SPEECH RECOGNITION IN NOISE USING MODEL BASED ADAPTATION STRATEGIES
M. Kadirkamanathan · 2024
The performance of misting speech recognition systems which are designed to operated in low noise background noise environments have been known to degrade quite signi cantly with increasing noise levels [6].Among the many techniques proposed for noise robustness in recognisers, those based on adapting speech models trained in one environment to the ambient conditions have shown much success.The Hidden Markov Model (rm) decomposition reoognitionlll and the model oomhinationlsl technique have illumted good performances in rather meme conditions.The Klatt noise masking ] is another algorithm of this type.Acoustic ambient noise is usually considered to be additive.The sampled signal is the sum of the acoustic speech signal and the acoustic ambient signal.The frontrend of most speech recognisers perform a short term spectrum analysis on the sampled signal as a rst step.Estimates ofthe signal energy in various frequency bands are calculated in dB or equivalent energy level measures.As~ suming that the moss correlation term between the speech signal and the ambient noise signal is negligible compared with the autoosrrelation terms, the ergy estimate may be elpressed as