A method for detecting a transient in an audio signal that has been broken up into frames includes obtaining a time domain feature of the frames and comparing the domain feature with a predetermined value. If the time domain feature is greater than the predetermined value, the frames are taken as transient and if the time domain feature is less than the predetermined value, the frames are taken as non-transient. The method has a low computational intensity and is thus very suitable for devices with limited processing resources.

Method for detecting audio signal transient and time-scale modification based on same

Diphone synthesis has been first introduced in France by LEIPP and al. in 1967 in relation with a perceptive theory describing the speech structures. An intelligible, monotonous voice was synthesized by means of a 44 oscillator device named ICOPHONE, from a lexicon of some 600 normalized diphones. As a result, the autonomous vocal response unit ICOPHONE 5, operational since 1974, produces fluent French in real-time from the text written in orthographic or phonetic form. Vocal response, including prosody, should be entirely automatic, and work even with non-grammatical sentences. An algorithm has been written with respect to these considerations : pitch and duration are deduced from the text, without any syntax analysis or manual marking. Present results confirm the idea that syntax is not the essential factor governing prosodic contours.

Diphone synthesis of French: Vocal response unit and automatic prosody from the text.

Tracking and identifying the formants in order to perform speech recognition is a time-consuming, error full and speaker-dependent operation. It is proposed to characterize the speech short-term spectrum with a reduced number of parameters (4 to 7) computed from a rough spectral analysis. These parameters permit a correct classification of the steady-state French speech sounds (vowels, including nasals, and unvoiced fricatives) pronounced by different speakers. A word recognition experiment based on the same parameters gives good results with words differing from each other by one phoneme only (single speaker, one learning pass).

Speech characterization from a rough spectral analysis

Cette these presente un traitement inspire du fonctionnement du systeme auditif pour ameliorer la reconnaissance vocale. Pour y parvenir, le signal de la parole est filtre par un banc de filtres et compresse pour en produire une representation auditive. 
L'innovation de l'approche proposee se situe dans l'extraction des elements acoustiques (formants, transitions et onsets) a partir de la representation obtenue. En effet, une combinaison de detecteurs composes de neurones a decharges permet de reveler la presence de ces elements et genere ainsi une sequence d'evenements pour caracteriser le contenu du signal. 
Dans le but d'evaluer la performance du traitement presente, la sequence d'evenements est adaptee a un systeme de reconnaissance vocale conventionnel, pour une tâche de reconnaissance de chiffres isoles prononces en anglais. Pour ces tests, la sequence d'evenements agit alors comme une selection de trames automatique pour la generation des observations (coefficients cepstraux). En comparant les resultats de la reconnaissance du prototype et du systeme de reconnaissance original, on remarque que les deux systemes reconnaissent tres bien les chiffres prononces dans des conditions optimales et que le systeme original est legerement plus performant. Par contre, la difference observee au niveau des taux de reconnaissance diminue lorsqu'une reverberation vient affecter les donnees a reconnaitre et les performances de l'approche proposee parviennent a depasser celles du systeme de reference. De plus, la selection de trames automatique offre de meilleures performances dans des conditions bruitees. 
Enfin, l'approche proposee se base sur des caracteristiques dans le temps en fonction de la nature du signal, permet une selection plus intelligente des donnees qui se traduit en une parcimonie temporelle, presente un potentiel fort interessant pour la reconnaissance vocale sous conditions adverses et utilise une detection des caracteristiques qui peut etre utilisee comme sequence d'impulsions compatible avec les reseaux de neurones a decharges. 
Mots-cles : Intelligence artificielle, reconnaissance vocale, extraction de caracteristiques phonetiques, systeme auditif biologique, approche connectioniste, modeles de Markov caches.

Traitement bio-inspire de la parole pour systeme de reconnaissance vocale

We present here several points concerning speech communication with computers, and dealing with different aspects of the problem. We do not intend to write about a definite part of a speech communication system, but rather to give in formation and to report experiments that may be useful for people working in that field.

Some Points Concerning Speech Communication with Computers

Acoustic-phonetic conversion is probably the most critical step in continuous speech recognition. The transitional information can be used as follows, in order to improve the results. First we contitute a lexicon of the phoneme steady-state spectra and a lexicon of all the transitions (diphones), each one being characterized by a"differential spectrum". The unknown continuous speech wave is segmented into quasi steady-state and transitional segments ; the labelling of the quasi steady-state segments admits several candidates. The transitional segment between two quasi steady-state spectra is then compared to the diphones of the lexicon selected from the combination of the surrounding possible phoneme labels. Actually, only the comparisons which are compatible with the recent past of the message are made. When working as a phoneme-vocoder, the whole procedure needs about 3x real-time, without any optimization.

http://ieeexplore.ieee.org/iel6/8365/26347/01170171.pdf

J. Lienard

Papers

Acoustic-phonetic recognition of connected speech using transient information