Audio Documents Restoration as a Documentary Source in the Linguistic Research: Comparison of Instruments

In the last years, always more speech audio archives digitized their historic corpora with the aim of preserving them and exploiting the advantages offered by digital signal processing techniques, such as formant analysis, automatic transcription, content-based information retrieval. Many of these applications are less effective when the signal to noise ratio (SNR) decreases, as it often happens in field-recorded linguistic corpora, due to the use of non-professional portable analogic devices and environmental noises. Audio restoration tools can therefore be very useful to increase SNR and improve subsequent analyses. At the same time, restoration algorithms should be carefully chosen and tuned to avoid the distortion of essential characteristics, such as formant position and energy. The paper describes, with examples taken from on field-recorded sound excerpts, several algorithms, able to cover different audio restoration categories. The algorithms were then evaluated by means of an automatic speech recognition tool. Results showed that the noise reduction algorithms can improve the phoneme recognition task carried out with the Kaldi toolkit.