Improved Jacobian adaptation for fast acoustic model adaptation in noisy speech recognition

This paper describes two algorithms to improve a previously proposed Jacobian adaptation (JA) technique for fast acoustic speech recognizer model adaptation in environmental noise. The rst technique introduces a new bias term, that is a function of the reference noise estimate to account for the mismatch between the reference noise estimate and noise component of the noisy speech spectrum. This functional mismatch bias is quite general, and here we choose to represent it as a linear function of the reference noise estimate. The second algorithm uses a more accurate relationship between the log and linear spectral domain versions of the HMM parameters. The combination of these new techniques achieves an increase of between 3.3-10.9% in recognition accuracy in adapting from automobile highway noise (HWY1) to such low-frequency noise sources as IBM PC cooling fan noise (PS2), large city street noise (LCI) and HWY2 (a di erent highway noise with di erent automobile at di erent signal-to-noise ratio (SNR)) over the original Jacobian adaptation in context independent phone recognition on TIMIT database.