Towards efficient and scalable speech compression schemes for robust speech recognition applications

Efficient scalable encoding for distributed speech recognition

“…6 and 7 is that the bitrate required to ensure that speech recognition performance is not degraded due to compression is about 4600 b/s for the spoken names task and 5700 b/s for the CSR task. Additionally, it has been shown that the approximate minimum bitrate for transparent operation for an isolated digits task was 1100 b/s (Srinivasamurthy et al, 2001a) and for a connected digits task was 2000 b/s (Srinivasamurthy et al, 2001b). This illustrates that the minimum bitrate for transparent speech recognition is strongly task dependent.…”

“…Due to this overlap and the underlying correlation in the speech (because of the slow movement of articulators), it is reasonable to expect that MFCC vectors from adjacent frames will exhibit high correlation. To achieve good compression efficiency this correlation has been exploited using linear prediction (Ramaswamy and Gopalakrishnan, 1998;Srinivasamurthy et al, 2000), where a given MFCC in a frame was predicted from the corresponding MFCC in the previous frame. 2 The prediction error e i ¼ u i À aû iÀ1 was quantized using uniform scalar quantization (USQ), where u i is the current sample andû iÀ1 is the reconstruction of the previous sample generated by the coarse prediction loop.…”

“…The effect of various speech coding techniques on speech recognition, including GSM (Digalakis et al, 1999;Srinivasamurthy et al, 2000;Kiss, 2000;Lilly and Paliwal, 1996;Srinivasamurthy et al, 2001b), G.723.1, G.727, G.728, G.729 (Turunen and Vlaj, 2001) (Lilly and Paliwal, 1996) and MELP (Srinivasamurthy et al, 2000;Srinivasamurthy et al, 2001b), has been previously evaluated by a number of researchers. In all cases, it was shown that speech coding significantly degrades speech recognition performance.…”

Signal Processing Challenges in Distributed Stream Processing Systems

“…Another approach consists in providing each processing task with the most relevant information in order to maximize its classification accuracy. One potential solution to this problem has been addressed in [26] in the context of scalable speech recognition. The authors considered two sequential speech recognition systems with very different resource requirements.…”