This paper proposes an algorithm for recognizing noisy-speech while avoiding the tedious training of noisy-speech HMMs. HMM composition combines a noise-source HMM and a phoneme HMM into one noise-added phoneme HMM. The speech recognizer is based on LPC cepstrum analysis. In the first set of speaker-dependent experiments consisting in recognizing 23 Japanese phonemes with a variety of stationary and nonstationary noises with signal-to-noise ratios ranging from 0 dB to 20 dB, the algorithm reduced the error of the phoneme-recognition rate by more than 75%. In the second set of speaker-dependent experiments consisting in recognizing continuous speech sentences, the composed HMMs could be obtained very rapidly and gave similar recognition rates to those of phoneme HMM models trained by using a large noise-added speech database. The efficiency, flexibility of the algorithm and its adaptability to new noises and to various SNRs make it a suitable basis for a real-time speech recognizer resistant to noise.
Keywords: noisy speech, HMM composition