An Ultra-Low Bitrate Neural Audio Codec Under NB-IoT

With the advancement of NB-IoT, the demand for speech communication has surged in applications such as medical assistance, smart alarms, and smart homes. Consequently, transmitting high-quality data with limited speech bandwidth has become a pressing concern. Traditional ultra-low bitrate speech coding techniques often fail to capture the long-term dependencies of time-series data, resulting in poor quality at lower data rates. To fully explore the correlation between data, this paper introduces an interleaved structure comprising dilated convolutions and Intra-LSTM. Experimental results demonstrate the effectiveness of this structure in enhancing both intra-frame and inter-frame correlations. Additionally, to address the issue of inattentiveness within the interleaved structure, Attention gates are introduced to effectively capture the periodicity and similarity of input sequences. Finally, this paper presents an end-to-end neural speech codec capable of operating at an ultra-low bitrate of 0.5kbps. Experimental findings demonstrate that our proposed codec significantly outperforms traditional speech codecs in terms of reconstructed speech quality at 0.5kbps, achieving a PESQ (Perceptual Evaluation of Speech Quality) score of 1.86, representing a 25% improvement over the Speex method. Moreover, subjective experimental results indicate that the reconstructed speech quality is comparable to that of Lyra at 3.2kbps. This research holds significant practical implications for advancing narrowband IoT communication systems.