back to DITELO
 
Data collected on Field

During the test phase of the automatic services 170 and Italy Direct, we have collected a number of field signals. Our achitecture (pure software) allows to store on disk exactly the signal that will be recognized, so data collection is quite cheap. Data check, on the other hand, is quite expensive. These data are part of a new database, called FIELD, that collects digit sequences in continuous speech and confirmations. Most of the signals are corrupted by noise (cars, background speech, tv, music, babies, etc.) and contain weak spontaneous speech phenomena (hesitations, breaths, coughs, etc.). False starts are unfrequent, while often out-of-vocabulary words are present.

Some samples of the database FIELD show why recognition is difficult:

speech signal is often noisy, as many calls come from public boxes (background voices and music);

zero sei nove nove nove sei due quattro due
spectrogram (110 kByte)
file audio .riff (112 kByte)
file audio sphere (58 kByte)

especially users B (which sometimes do not immediately realize to speak with an automatic system) tend to give out-of-vocabulary answers ("OK", "yes, I do accept the call from Germany", etc.). This phenomenon can be reduced, but not eliminated, by choosing appropriate prompt messages.

BOCCA si' accetto RESPIRO
spectrogram (88 kByte)
file audio .riff (78 kByte)
file audio sphere (21 kByte)

as calls are from/to abroad, often some tones are added to the signal;

TU TU si' RESPIRO TU TU
spectrogram (91 kByte)
file audio .riff (32 kByte)
file audio sphere (11 kByte)

weak spontaneous speech phenomena (hesitations, breaths, coughs, etc.) are frequent;

EH si'
spectrogram (82 kByte)
file audio .riff (36 kByte)
file audio sphere (12 kByte)

  Last update 10/3/1998 - Maintainer Roberto Gretter