Method of speech recognition with learning

A speech processing method comprising a learning phase during which a signal originating from a sensor is processed to obtain parameters which represent voice structures uttered by a user in a mildly noisy environment and said parameters are stored in a dictionary, and a subsequent recognition phase during which the signal originating from the sensor is processed to obtain parameters which are compared with those stored in the dictionary so as to recognize voice structures uttered by the user in a noisy environment, wherein the obtaining of said parameters during the learning and recognition phases includes the steps of:

• forming frame of predetermined length from the signal originating from the sensor;
• transforming each frame from the time domain to the frequency domain to obtain spectrum X(i);
• and
• determining the magnitude .vertline.X(i).vertline..sup..gamma., where .vertline.X(i).vertline. represents the magnitude of the spectrum and .gamma. represents an exponent lying between 0 and 2,
• applying an inverse transformation, from the frequency domain to the time domain, to said magnitude .vertline.X(i).vertline..sup..gamma. ;
• and
• wherein, in the recognition phase, the exponent .gamma. is adapted dynamically for each frame.

A speech processing method comprising a learning phase during which a signal originating from a sensor is processed to obtain parameters which represent voice structures uttered by a user in a mildly noisy environment and said parameters are stored in a dictionary, and a subsequent recognition phase during which the signal originating from the sensor is processed to obtain parameters which are compared with those stored in the dictionary so as to recognize voice structures uttered by the user in a noisy environment, wherein the obtaining of said parameters during the learning and recognition phases includes the steps of:

• forming frames of predetermined length from the signal originating from the sensor;
• transforming each frame from the time domain to the frequency domain to obtain a spectrum X(i);
• and
• determining the magnitude .vertline.X(i).vertline..sup..gamma., where .vertline.X(i).vertline. represents the modulus of the spectrum and .gamma. represents an exponent lying between 0 and 2;
• applying an inverse transformation, from the frequency domain to the time domain, to said magnitude .vertline.X(i).vertline..gamma.;
• and
• wherein, in the recognition phase, the exponent .gamma. is selected from among a plurality of discrete values lying between 0 and 2.