Speech recognition system and method

A speech recognition method, comprising:

• determining a speed level of a moving object by using a noise signal at an initial time of speech recognition;
• enhancing, if the speed level is lower than a specific level, sound quality of an input speech signal by extracting voice activity detection features robust against vehicular noise and applying to the input speech signal a Wiener filter using the extracted voice activity detection features;
• detecting start and end points of a signal having been subjected to the sound quality enhancement, if the speed level is lower than a specific level;
• eliminating sudden noise components occurring when the moving object moves based on a sudden noise Gaussian mixture model, if the speed level is lower than a specific level;
• and
• decoding a signal having been subjected to the sudden noise elimination by using a low-speed/medium-speed acoustic model database, if the speed level is lower than a specific level.

18 The speech recognition method of claim 13, further comprising:

• enhancing the sound quality of the input speech signal of the speech recognition by applying a Wiener filter using a Gaussian mixture model, if the speed level of the moving object is higher than the specific level;
• detecting start and end points of a signal having been subjected to the sound quality enhancement, if the speed level is higher than the specific level;
• compensating distorted harmonic components of the signal having been subjected to end point detection based on a harmonic Gaussian mixture model, if the speed level of the moving object is higher than the specific level;
• eliminating sudden noise components occurring when the moving object moves based on a sudden noise Gaussian mixture model, if the speed level of the moving object is higher than the specific level;
• and
• decoding a signal having been subjected to the compensation and the sudden noise elimination by using a high-speed acoustic model database, if the speed level of the moving object is higher than the specific level.

The speech recognition method of claim 18, wherein said enhancing the sound quality if the speed level is higher than the specific level includes:

• estimating a de-noised power spectral density of the input speech signal;
• calculating a zero-mean power spectral density by using the de-noised power spectral density;
• calculating a first clean speech power spectral density by using a power spectral density of clean speech and the zero-mean power spectral density;
• calculating a second clean speech power spectral density by using the de-noised power spectral density and the first clean speech power spectral density;
• acquiring frequency response characteristics of the Wiener filter using the Gaussian mixture model by using the second clean speech power spectral density;
• and
• enhancing the sound quality of the input speech signal based on the frequency response characteristics.