The HMMs are widely-used statistical models to characterize the sequence of speech spectra and have successfully been applied to speech recognition systems. This paper discusses a hidden Markov model (HMM) based on multi-space probability distribution (MSD). Our experimental analysis and results support us to conclude that the accent is a serious problem in Mandarin speech recognition and the MLLR adaptation is effective in reducing the mismatch caused by accents. Experimental results show that up to 45% of the accent related errors get corrected for accented speakers and there is no such improvement for standard speakers. For the second question, we perform maximum likelihood linear regression (MLLR) adaptation for each speaker and then analyze the recognition results. We analyze their effects on speech recognition for each speaker. For the first question, we focus on three types of mispronunciations as the dominant problems of accented Mandarin. What we want to address in this paper are two questions about Mandarin: whether accents affect speech recognition greatly how we can solve the problem. Across a vast region and a huge population, there are varieties of accented Mandarin spoken in China, which are mainly caused by speakers' dialects.

This paper presents our work on accent issues in Mandarin large vocabulary continuous speech recog-nition. The results of these experiments demonstrate that the proposed Full-band LPCNet is the only neural vocoder that can synthesize high-quality 48 kHz speech waveforms while maintaining real-time capability with a CPU. We then evaluate the synthesis quality using both normal speech and a singing voice. In this paper, we propose a Full-band LPCNet to synthesize high-fidelity 48 kHz speech waveforms with a CPU by introducing some simple but effective modifications to the conventional LPCNet. LPCNet was proposed as a real-time neural vocoder with a mobile CPU but its sampling frequency is still only 16 kHz. Although most previous studies on 48 kHz speech synthesis have used traditional source-filter vocoders or a WaveNet vocoder for waveform generation, they have some drawbacks regarding synthesis quality or inference speed. This paper investigates a real-time neural speech synthesis system on CPUs that can synthesize high-fidelity 48 kHz speech waveforms to cover the entire frequency range audible by human beings.

Open source speech to text