The Frontend Noise Reduce Of Isolated Word Speech Recognition

The object that speech recognition technology researches is speech signal processing. This technology can make the machine understand natural language of mankind and can be human-computer interaction interface.Now, in ideal environment, most speech recognition products can have a high recognition rate. But in noisy environment, the rate reduced rapidly. So how to improve the performance of the recognition system in noisy environment is very significant and the market prospect is very broad. The content of this paper mainly research the frontend noise reduce of speech signal which is applied in Chinese speech recognition chip of small-vocabulary, speaker-independent isolated word. The main work is: studying the noise reduction algorithms and the feature extraction of different sampling rate speech signal.The algorithm which is researched in this paper is based on Wiener filter, spectrum estimation and voice activity detector. Speech features are computed from the input signal in the Feature Extraction part. Then, features are compressed and further processed.In the Feature Extraction part, waveform processing is applied to the de-noised signal and cepstral features are calculated. At the end, blind equalization is applied to the cepstral features. The Feature Extraction part also contains 16 kHz extension block for handling these sampling frequencies. Voice activity detection (VAD) for the non-speech frame dropping is also implemented in Feature Extraction.By the above process of noise reduction and feature extraction, the floating-point recognition rate of isolated words reached 95%, the robustness of recognition has been greatly improved. Then the floating-point algorithm is converted to the fixed-point.