Imperfect transcript driven speech recognition

HAL Id: hal-02094739

https://hal.archives-ouvertes.fr/hal-02094739

Submitted on 9 Apr 2019

HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

Imperfect transcript driven speech recognition

Benjamin Lecouteux, Georges Linares, Pascal Nocera, Jean-François Bonastre

To cite this version:

Benjamin Lecouteux, Georges Linares, Pascal Nocera, Jean-François Bonastre. Imperfect transcript driven speech recognition. Interspeech, 2006, Pittsburgh, United States. �hal-02094739�

Im pe rfe ct transcript drive n spe e ch re cognition

Be njam in Le coute ux, Ge orge s Linarès, Pascal Nocéra, Je an-François Bonastre

Unive rsité d'Avignon - {b e njam in.l e coute ux, ge orge s.l inare s, pascal .noce ra, je an-francois.b onastre }@ univ-avignon.fr

- actor

- journal ist

Spe e ch re cognize r (b ase d on A* al gorith m )

- prom pte r te xts - sce nario scripts

Autom atic transcript al igne d and corre cte d

Approxim ate d transcript (~ 10% W ER)

LM traine d on im pe rfe ct transcript Ge ne ric

l anguage m ode l (65K w ords)

70% 30%

Inte rpol ate d l anguage m ode l

audio stre am

Autom ate d transcript al igne d and corre cte d

Audio stre am

Re su l ts :

- Initial de coding : 22,7% W ER using a ge ne ric l anguage m ode l .

- D e coding using l anguage m ode l inte rpol ation.

- D e coding using LM

inte rpol ation and al ignm e nt.

M L-G : Ge ne ric m ode l l anguage 65K w ords

M L-TrErr : Language m ode l traine d on th e transcript al TrEr : Al ignm e nt to th e im pe rfe ct transcript

Expe rim e ntal conte xt :

- Expe rim e nts asse sse d on 3 h ours of radio broadcast - 10 % W ER introduce d in transcripts

- Ge ne ric l anguage m ode l use d : 65000 w ords traine d on "Le M onde "

- Spe e ch re cognition syste m : SPEERAL, an

asynch ronous de code r base d on th e A* al gorith m .

Al ignm e nt w ith A* al gorith m d u ring asynch ronou s d e cod ing

- Each e val uate d w ord is al igne d to th e re fe re nce w ord stre am using a D ynam ic Tim e W arping (D TW ) al gorith m .

- O nce th e h ypoth e sis is synch ronize d w ith th e transcript, th e al gorith m e stim ate s th e m atch ing transcript-to-h ypoth e sis score .

- Th e n, l inguistic probabil itie s are m odifie d using th e fol l ow ing re scoring rul e :

- is m axim um w h e n th e trigram is al igne d and de cre ase s according to th e m isal ignm e nts of th e h istory.

Al ignm e nt e xam pl e .

Prom pte r : france inte r fl ash d'inform ations à h uit h e ure s un q uart Pronounce d te xt : france inte r l ' actual ité à h uit h e ure s un q uart

Re su l ts afte r d e cod ing :

W ith out al ignm e nt : FACE À ELLE actual ité à h uit h e ure s ET q uart W ith al ignm e nt : france inte r l ' actual ité à h uit h e ure s un q uart

O n-d e m and synch ronization and e stim ate fu nction re scoring

O n-de m and synch ronization of th e A* se arch w ith im pe rfe ct transcript b y th e D TW al gorith m

al ignm e nt al ignm e nt

W ER

Concl u sions :

- Be st re sul ts : inte rpol ation 70-30% w ith force d al ignm e nt to transcript

- W ER be tte r th an th e transcript : 7.2%

- re l ative im prove m e nt : 28%