Enhancing speech intelligibility in noise

We use a very simple method to modify the speech signal (while preserving its energy!!) in such manner that it is more intelligible when mixed with noise at very low SNR. In this example, speech-shaped noise at -9dB SNR is used. The method involves two steps: spectral tilt modification and dynamic range compression. These operations are carried out on the parameters of a harmonic model.

Original speech:

Original speech + noise:

Modified speech + noise:

We have also applied these findings to the development of a noise-robust HMM-based TTS system where signal modifications are embedded in the generation process.

Original synthetic speech:

Original synthetic speech + noise:

Modified synthetic speech + noise:

[Back]