FR2851352A1 - Continuous audio signal converting system for use in linguistic converter, has vocal synthesizer that synthesizes translated textual portions into synthesized portions that are mixed with respective residual portions by mixer - Google Patents

Abstract

The system has a vocal recognition module (6) converting vocal portions into textual portions of a language. The textual portions are translated into translated textual portions of another language by a translation module (7). A vocal synthesizer (8) synthesizes the translated textual portions into synthesized portions that are mixed with respective residual portions, to obtain a converted audio signal by a mixer (9). The vocal portions (PT) and the residual portions (PR) of same duration are formed by segmenting a received audio signal (SA) into audio portions (PAC) including the vocal and residual portions, by an audio analyzer.

Description

System for converting a continuous audio signal into a

  Translated and synthesized audio signal The present invention relates to a system for converting a continuous audio signal of indefinite duration, containing vocal portions in a first language, into an audio signal translated and synthesized in a second language different from the first language.

  Currently, some systems translate audio signals in real time. According to European patent application EP 1093059, a translation system is composed of a voice recognition unit, a translation engine, and a voice synthesis unit. The voice recognition unit continuously analyzes an incoming audio signal in a first language, and gradually submits a text result to the translation engine. The translation engine 20 continuously generates a text translation into a second language from the result of the voice recognition. The translation is provided to the text-to-speech unit. Based on a comparison between the current translation and the previous 25 translations, the text-to-speech unit synthesizes the result of the translation into the second language.

  Recent research has improved the quality of translations. However, current translation systems do not distinguish between the actual voice signal and a residual signal or background noise in the incoming audio signal, or portions of the incoming audio signal which are not to be translated. These distinctions generate a temporal problem posed by an incoming audio signal of indefinite duration at the level of the processing time by the system and more precisely different translation times according to the language used and according to the portions of the incoming audio signal to be translated. They cannot therefore incorporate new techniques such as taking into account in the translation the context in which the words of the audio signal to be processed are situated.

  The objective of the present invention is to translate and synthesize a continuous audio signal of indefinite duration by overcoming any time constraint due to the duration of processing by the conversion system and to the variations in the durations of the translated portions of the outgoing continuous signal. compared to the durations of the corresponding portions initially not translated in the incoming audio signal.

  To achieve this objective, a system for converting a continuously incoming audio signal including a voice signal in a first language into an audio signal converted in a second language, comprises means for segmenting the received audio signal into audio portions, means for filtering the audio portions each into a voice portion and a residual portion of the same duration, voice recognition means converting the voice portions into text portions in the first language, means for translating the text portions into translated text portions in the second language, and means for vocally synthesizing the textual portions translated into synthesized portions having different durations respectively from the audio portions. The system is characterized in that it comprises a mixing means for transforming each residual portion and the respective synthesized portion into portions which have a duration less than the longest of the durations of the residual portion and of the respective synthesized portion and 5 which are in phase in order to mix them into a mixed portion making up the converted audio signal.

  The invention also remedies any delay or advance too substantial of one of the converted audio signal and incoming audio signal relative to the other in order to smooth the time differences between these two signals. To this end, the system may include means for detecting a temporal excess of a sum of the durations of the portions of the converted signal over a sum of the durations of the corresponding portions of the incoming audio signal so that the means for segmenting deletes a portion of the incoming audio signal whose duration is substantially equal to said time surplus, and / or a means for detecting a time deficit of a sum of the durations of the portions of the converted signal with respect to a sum of the durations of the corresponding portions of the incoming audio signal so that the means for detecting adds an audio portion to the converted signal, the duration of which is substantially equal to said time deficit. Other characteristics and advantages of the present invention will appear more clearly on reading the following description of several preferred embodiments of the invention with reference to the corresponding appended drawings in which: - Figure 1 is a schematic block diagram of a translation and speech synthesis system according to a preferred embodiment of the invention, in the environment of a terminal user installation comprising several receiving devices and several processing servers according to the invention; FIG. 2 is an algorithm of steps executed by the system according to the preferred embodiment for converting a continuous audio signal into a translated and synthesized audio signal; and - Figure 3 is a schematic block diagram of a preferred embodiment of a language converter included in the conversion system 10 according to the invention.

  In the following, the term "chain" denotes either a channel or a transmission channel for broadcasting a sound broadcasting program 15 and the program company broadcasting said program.

  The term "program" designates a succession of sound broadcasting programs, also called magazines, broadcast by a specific channel.

  With reference to FIG. 1, the system for converting a continuous audio signal including a voice signal into an audio signal by translation and speech synthesis according to a first embodiment of the invention essentially comprises an IT user terminal installation and a SE audio conversion server, or more generally several translation servers.

  The IT user terminal installation includes M 30 receiver equipment EQ1, ... EQm, ... EQM with 1 <m <M. For example, one EQ1 of the devices is a sound broadcasting receiver capable of receiving selectively the broadcasts of several sound broadcasting stations (stations). Another EQM equipment is a personal computer (PC), for example connected to a packet network of the Internet network type, or connected to a wired network for the distribution of sound broadcasting programs.

  The equipment EQ1 to EQM is controlled through a distributed bus BU by a central processing unit UCit in the IT installation. As a variant, all or part of the BU bus can be replaced by a proximity radio link of the Bluetooth type or according to the 802.11b standard.

  The UCit central unit essentially comprises a microcontroller connected to a communication interface IC and optionally to a keyboard and a screen. The central unit and the communication interface are physically included in a housing independent of the equipment. As a variant, the UCit central unit with the peripherals is integrated into the computer or the broadcasting receiver or the television receiver EQm.

  In another variant, the terminal installation is reduced to a simple receiver. The UCit central unit constitutes a basic module which can serve various home automation equipment such as that illustrated in FIG. 1 as well as one or more mobile telephones and radiotelephones, at least one television receiver, an alarm central, etc. . The communication interface IC is adapted to a telecommunications link LT connected to an access network RA of the installation IT. The link LT and the network RA can conventionally be a telephone line and the switched telephone network PSTN itself connected to a high speed packet transmission network RP of the internet type. According to other variants, the telecommunications link LT is an xDSL line (Digital Subscriber Line) or an ISDN line (Digital Network with Integration of Services) connected to the corresponding access network. The link LT can also be confused with one of the links serving one of the equipment's EQm through one of the distribution networks RD defined below.

  Figure 1 also schematically shows the telecommunications system surrounding the IT user terminal installation. In particular, the references RD and TR denote respectively one or more distribution networks for scheduled sound and television broadcasting programs and one or more head ends broadcasting programs and managed by various sound broadcasting and program companies. of TV. All of the RD distribution networks include in particular analog and / or digital broadcasting networks for broadcasting programs capable of being received at least by the radio receiver EQ1. The set of distribution networks RD also includes the Internet network through which the computer EQm is capable of receiving radio broadcasts broadcast by certain program companies.

  Each audio conversion server SE is connected to the program distribution network RD and to the terminal installation of the user IT via the packet network RP and the access network RA. According to another variant, the functionalities of the conversion server SE are located in a headend TR, or more generally, the server SE is connected to the broadcast distribution networks RD. In this case, the translation and the speech synthesis are carried out before broadcasting.

  The scheduled programs, except those live, are translated and synthesized by slight anticipation, at least a few minutes before their broadcast, which offers translation and voice synthesis 5 almost without time difference. Indeed, as explained below, the conversion of an audio signal SA including a voice signal into a first language, such as English, by the conversion server has a certain duration which generates a time delay between the audio signal entering SA into the system and the converted audio signal SAC leaving the system and having a voice signal in a second language, such as French. When the conversion occurs when listening to a continuous audio signal, the variable delay d to the conversion can be made up by the continuous audio signal which will then be duplicated but translated at the start of the processing, or by an audible message of the type "translation in progress", or simply by silence, or by any other predetermined audio sequence.

  The server SE comprises a central processing unit UCs and a set of peripherals including at least one database BD and a linguistic converter CL described in detail below.

  Many variants of the hardware distribution of the components of the IT user terminal installation and of the processing server ST can be deduced from the embodiment of the invention illustrated in the figures.

  According to a variant of architecture called "thick client / thin server", as opposed to the preferred embodiment illustrated in FIG. 1 called "thin client / heavy server", the conversion system is partially included in the terminal user installation IT. The linguistic converter CL and the database BD are installed in the user installation IT, and the processing which was carried out by the central unit UCs is then executed in the processing unit UCit.

  Other intermediate variants between the thin client / heavy server architecture and the thick client / thin server architecture are conceivable. For example, modules making up the language converter CL are distributed in the terminal installation of user IT and the server SE.

  According to another embodiment, all of the processing carried out subsequently is executed upstream of the broadcasting of the programs, in a network head TR. In this case, the user's terminal installation is almost reduced to equipment EQ1 to EQM.

  The terms "conversion parameters" refer to PAC activation parameters, an IL language identifier and general PAG parameters. The activation parameters characterize an activation period of the conversion system according to the invention as a function of start and end dates and times and / or of the type of program.

  The language identifier IL defines a translation language. The general parameters define a list of user preferences such as the possibility of translating or not translating advertisements and / or songs contained in the audio signal SA.

  In another embodiment of the invention, a preference program is used to store in the database BD and set preferences on the translation and the speech synthesis desired by the user in order to establish and store the conversion parameters and modify them if desired.

  The preference program is executed by the server SE via the packet network RP, or directly by the central unit UCit of the terminal installation IT when the database BD is included in the installation IT.

  For example, the preference program 10 presents a complete list of the equipment EQ1 to EQM of the user via a display in the IT installation, for example the screen of the personal computer EQM, so that the user selects the equipment for which he wishes to modify audio conversion parameters when the identifiers of several user equipment have been registered during his subscription. Audio conversion parameters can be proposed to the user by default, or the current parameters if the user has already selected or modified these parameters. A first page invites the user to enter PAC activation parameters programmable by the user according to dates and times or directly according to programs chosen from a program schedule. Each time the user validates an entry page, the entered values of the parameters are sent to the ST server for storage in the BD database, or directly in the BD database of the terminal installation for architecture. 30 "thick client / thin server". The same is true for IL language identifiers.

  As a variant, linguistic parameters PL are added to the conversion parameters and characterize the sound presentation of the one or more voices synthesized in the outgoing voice signal SAC. Each voice depends on gender, age group, prosodic characters, accent, rhythm, etc. chosen by the user. As a variant, the linguistic parameters are those of at least one chosen from several predetermined voices. According to a simpler variant, the synthesized voice substantially reproduces the voice or the voices of several speakers in the incoming audio signal.

  If the IT terminal installation does not have a human machine interface such as a mouse or keyboard, the parameters corresponding to the user's preferences are selected by default.

  If the audio conversion according to the invention is carried out in a headend TR and the terminal installation IT is reduced essentially to the equipment EQ1 to EQM, the parameters are modified by the user via any other means, for example by a terminal by telephone or radiotelephone or by an operator when subscribing to the audio conversion service according to the invention.

  FIG. 2 shows an algorithm of steps E1 to E8 executed by the conversion system according to the preferred embodiment for converting a continuous audio signal SA transmitted by the distribution network RD to one EQm of the receiving equipment of the IT installation into a converted SAC audio signal.

  In step E1, the user U of the IT installation 30 powers it up and selects an equipment EQm in order to globally activate the conversion system of the invention. For example, a predetermined pressure of a remote control of the selected equipment EQm when this selected equipment contains the central unit UCit, or a switch to the start position of a button on the housing incorporating the central unit UCit powers up the UCit unit. This reads from memory and then automatically transmits an identifier IU of the user U and an identifier IEQm of the equipment EQm selected by the user U to the server SE.

  The server SE identifies the user U who has subscribed to the audio conversion service, by comparing the identifier received IU with the identifiers of the users subscribed in the database BD, in step E2. In a variant, the server SE requests the user to enter in the IT installation the identifier UI and a password which has been assigned to him when subscribing to the service in order to transmit the identifier and the word to the SE server for verification. Then in step E2, the central unit UCs reads the conversion parameters PAC, IL, PAG and possibly PL from the database BD in correspondence with the user identifier IU in order to analyze them according to the following steps in order to produce the conversion in the selected equipment EQm for the selected chain. The PAC activation parameters are considered by the central unit UCs, so that the system is active only during the activation time determined by the PAC parameters.

  After the identification of the user in step E2, the central unit UCs in the server SE invites the user U to select a chain in the equipment EQm which then transmits an identifier ICH of the selected chain to the server SE via the UCit unit, in step E3.

  As a variant, the equipment EQm and the chain of the audio signal to be converted have been preselected by the user U in particular when subscribing to the audio conversion service, and the identifiers IEQm and ICH have been registered in correspondence with the identifier UI of user U in the database BD. In this variant, the EQm equipment is simply powered up awaiting an audio conversion.

  Alternatively when the receiving equipment is a personal computer connected to the Internet, the selection of a sound broadcasting channel is equivalent to calling up a website managed by the channel and designated by a precise URL address. In this case, after identification of the user, the user enters his parameters via the HTML page form.

  In step E4, the central unit UCs checks whether the incoming audio signal SA identified by the chain identifier ICH is being converted by the server SE and whether one of the 20 conversion parameters for the current conversion corresponds to one of the PAC, PAG and IL parameters selected by the user. When the conversion parameters correspond, the conversion is continued by step E7. Otherwise, the central unit UCs of the server SE selects the channel designated according to the identifier received ICH from all the channels available at the level of the server SE in step E5 in order to determine the incoming audio signal SA to be converted.

  In step E6, the linguistic converter CL filters the signal SA into voice portions PV and residual portions PB, transforms the voice portions PV of the audio signal SA into text portions PT, translates the text portions PT into text portions translated PTT according to the language identifier IL, synthesizes the textual portions translated into synthesized portions PS, respectively mixes the residual portions PB and the synthesized portions PS into 5 mixed portions PM, by temporally adapting the durations of the mixed portions to the durations of the original residual portions in the SA signal to finally compose an outgoing converted audio signal SAC. The operation of the converter is detailed later with reference to FIG. 3.

  The converted signal SAC is sent continuously during the progressive processing of the signal SA by the server SE to the terminal installation IT in step E7. All the conversion steps up to step E7 have caused a delay necessary for the execution of the audio conversion in the server SE.

  In step E8, the terminal installation IT receives the converted audio signal SAC for broadcasting on the bus BU to the selected equipment EQm so that the user U listens to the converted audio signal SAC in place of the initial audio signal SA corresponding to the ICH identifier of the selected channel.

  The conversion by the conversion system according to the invention is completed at the expiration of the activation time defined by the PAC activation parameters and monitored by the central unit UCs. The central unit UCs also monitors any modification of the conversion parameters in order to process it in real time. Similarly, when the user's terminal installation is stopped, the central unit UCit transmits an end of session signal to the central unit UCs.

  The conversion server SE comprises a linguistic converter CL, the operation of which is described below with reference to FIG. 3.

  The language converter according to the invention 5 comprises an audio analyzer 1, an audio database 2, an audio filter 3, a language determination unit 4, a voice analyzer 5, a voice recognition module 6, a translation module 7, a speech synthesizer 8, a mixer 10 9, a contextual database 10, a segment context determination unit 11 and a time adaptation unit 12.

  In the following, the term "context" designates a list of key words or expressions and their equivalents. Each key word or expression characterizes a context that can be addressed in any multimedia document. Certain contexts are combinations of contexts, or in the case of current or regional contexts, combinations of contexts specified by a proper name, such as for example: Brittany Weather, Afghanistan War, etc. It will be assumed that the incoming audio signal SA received by the server SE is digital; otherwise, the audio signal received is analog and converted by an analog-digital converter included in the audio analyzer 1.

  The audio database 2 stores pieces of audio data such as all or part of music, songs, jingles or commercials, news flashes and sound effects. More generally, the database 2 has previously recorded any piece of audio data preferably qualified by a type T, PAS audio parameters, duration and contexts CA whose time limits are staggered with respect to a fixed reference frame of audio data, such as the start of a song or a jingle, and possible translations or correspondences in specific languages and particularly in the second language corresponding to the identifier IL selected by the user U. The type T defines the nature of the piece of audio data in order to distinguish between them a song, an advertisement, a dialogue, a report, a particular theme, a sound effect, etc. The audio signal SA of the selected channel received by the selected equipment is temporarily stored temporarily in a buffer memory contained in the video analyzer 1. Like any audio signal SA, this includes periodic time marks such as words frame alignment, packet synchronization words, frame or line synchronization signals, etc. These time marks are counted modulo a predetermined number and stored in the buffer memory 20 contained in the video analyzer 1 in response to the selection of the ICH identifier of the channel by the user. These time marks are refreshed periodically during the processing of the audio signal SA by the system, in order to synchronize the processing of portions of audio signal by the various modules of the translation system.

  The audio analyzer 1 compares samples of the continuous incoming audio signal SA with samples of audio data contained in the audio database 2. Sequences of samples substantially similar to those in the audio database allow the analyzer audio 1 to delimit portions, called "known portions", included in the incoming audio signal SA corresponding to complete pieces or to portions of pieces of audio data contained in the audio database 2. The other portions of the audio signal SA which do not correspond to pieces of audio data or to parts thereof contained in the audio database 2, are called "portions to be converted" PAC. Among the known portions, the audio analyzer 1 decides as a function of the type T and of the duration of the current portion, of the conversion parameters chosen by the user, of the presence of a replacement portion associated with the known portion in the audio database 2, and the duration of said replacement portion to replace the known portion of the audio signal with the replacement portion. The replacement portion corresponds for example to the translation of the known portion into the second language chosen by the user. The known portions which have been replaced by their replacement portions are called "replaced portions" PRT, as opposed to the other known portions called "non-replaced portions" PNT. For example, the song-like portions may be non-replaced portions, and the advertisement-type portions may be replaced portions.

  Each portion is stored in the buffer memory, as well as its characteristics such as type T and its duration.

  The T type is applied to the speech synthesizer 8 so that it decides on a speech synthesis for the current portion to be converted as a function of the T type of the current portion.

  The non-replaced portions PNT and the replaced portions PRT are applied directly to the time adaptation unit 12 and also to the audio filter 3 followed by the voice recognition module 6 and 35 of the segment context determination unit 11 so that the contexts of these portions feed the contextual database 10. The portions to be converted PAC are transmitted only to the audio filter 3.

  In a variant, if the current known portion PNT or PRT has parameters PAS and a context CA read by correspondence in the audio database 2, the parameters PAS and the context CA are applied directly to the unit 11 in particular in order to 10 that they feed the contextual database 10 the corresponding portion is not transmitted to the audio filter 3. The audio analyzer 1 therefore participates in improving the quality of determination of the contexts, explained below, by directing the 15 PAS and CA parameters associated with the audio data and contained in the audio database 2 to the contextual database 10.

  Filter 3 filters by spectral subtraction or adaptive filtering the PAC or PNT or PRT portion of the audio signal SA applied by the audio analyzer 1 in order to dissociate it into a part comprising only voice and called "vocal portion" PV and a part residual PR including background noise and called "residual portion" PR. Filter 3 is for example based on a linear predictive analysis LPC (Linear Predictive Coding) and isolates different acoustic components in an audio signal such as voice, vocal noise and pure music.

  In order to improve the determination of the contexts, seen in detail below, the current residual portion PR comprising the residual non-vocal part of the current portion to be converted PAC produced by the filter 3 is returned by the filter 3 to the audio analyzer 1. The analyzer compares the residual portion with the portions of audio data contained in the audio database 2 and qualifies the residual portion PR by parameters PAS and a context CA of a portion of audio data similar or substantially identical to the current residual portion 5 PR. In order to rapidly constitute audio data in the database 2, the machines hosting the management means managing the audio database 2 can be shared. In another variant, the management means is associated with the audio analyzer 1.

  The current voice portion PV is processed in parallel by the voice analyzer 5 and the voice recognition module 6.

  The vocal analyzer 5 analyzes the current vocal portion PV 15 in order to determine the vocal parameters PVS characterizing sections allocated respectively to speakers and included in the vocal portion PV. The list of voice parameters is not fixed but includes, among other things, acoustic and particularly prosodic parameters such as the vibration frequency, intensity, flow, timbre and also other parameters such as the relative age of the speaker.

  In parallel with the voice analysis, the current voice portion 25 PV is subjected to the voice recognition module 6. When the language of the audio signal is considered to be unknown, the language determination unit 4 is inserted between the filter 2 and the voice recognition module 4. The unit 4 30 dynamically determines the language of the current voice portion PV of the audio signal SA if this is not previously known. For multi-language information for example, the language of the vocal portions is thus recognized continuously. If the language of the audio signal is predetermined and therefore that of the voice portion is predetermined, this language is taken as the default language and the language determining unit 4 is not necessary. The voice recognition module 6 transforms the 5 voice portion PV into a text portion PT.

  Several speech recognition modules can be used to optimize audio conversion. As a variant, the textual portion PT is subjected to a spelling corrector provided at the output of module 6.

  In another variant, the module 6 considers the results of a context study previously carried out by the segment context determination unit 11 in order to refine the voice recognition and the translation of the current voice portion PV. The context is translated into syntactic elements, that is to say key words and expressions, with high probabilities to be included in a portion of audio signal. 20 For example, the context of a relatively periodic or frequent commercial or news spot in an audio signal transmitted by a sound broadcasting station is predicted by knowing the detailed program of this station, or by deducing it from advertising spots or previous news.

  The context determination unit 11 further comprises a segmentation unit and a buffer memory. The segmentation unit receives the text portions PT produced by the module 6. The buffer memory continuously stores the current text portion PT for a duration DS less than the duration of a portion and greater than a duration making it possible to obtain a minimum number of words for contextual analysis, for example at least thirty words. The segmentation unit it segments the current text portion PT into temporal and periodic text segments ..., Sn, ... as and when the current text portion 5 is received. The predetermined duration of the audio portion segments depends on the relationship between the quality of analysis of the system, that is to say the relevance of the voice recognition as a function of the meaning of the words contained in the audio signal, and the time. processing by the system. For example, a segment duration of 20 seconds compared to a segment duration of 1 minute decreases the processing time by the system to the detriment of the quality of the voice recognition. A minimum duration of 15 to 15 seconds is typically sufficient for the system to ensure minimum quality.

  In another preferred embodiment of the invention, the segmentation is not based on a temporal characteristic but depends on a syntactic element such as a word, or a group of words or a sentence. A syntactic element is for example defined by a sound level above a predetermined threshold and framed by intervals of the audio signal having a sound level below the predetermined threshold and considered as silences.

  Various contexts in the form of key words and expressions, as defined above, deduced from segments preceding the current segment Sn and / or from the context study constitute contexts CSn 30 premoristered and managed in the contextual database 10 linked to the voice recognition module 6, to the translation module 7 and to the segment context determination unit 11. The contexts in the contextual database 10 35 are also completed and refined by automatic consultation of external contextual databases in depending on the contexts previously detected. The lists of contexts contained in these external databases are updated 5 manually and / or automatically. Each context is characterized by additional information such as the tone of the context, for example serious or happy, the segment of the population concerned, for example children, managers, workers ... The contexts 10 written in base 10 as and when their determination in unit 11 as well as the contexts contained in the external contextual databases are involved in the conversion of the voice portions PV into text portions PT 15 executed by the voice recognition module 6 and also in the translation of text portions in a translation module 7 described below. The contexts written in the contextual database 10 are thus progressively improved during the processing of the text portions PT to facilitate voice recognition in the voice recognition module 6 and translation in the translation module 7. The module 6 can '' press on Natural Language Understanding software 25 (Natural Language Understanding NLU).

  The unit 11 determines one or more contexts CSn of the current text segment Sn as a function of the current speech parameters PVS supplied by the analyzer 5 and as a function of the current text segment 30. In a preferred variant, contexts established and stored previously also serve for determining the context in unit 11 and contribute to increasing the relevance of new segment contexts which will in turn participate in determining contexts for next segments.

  In another variant, a general context is determined initially before any processing of the text portion PT as a function of parameters external to the system and linked inter alia to the source 5 of the audio signal SA. Program grids or information on the chain as well as any information likely to inform the context of the first textual segments of textual portion enrich the contextual database 10. 10 The unit 11 bases this general context on the context of a number determined from segments preceding the current segment Sn when the context of the immediately preceding segment is not determined and the preceding segment is the first segment of the current portion PT to be processed by the audio conversion system.

  As a variant, for the purpose of optimizing the conversion of a voice portion PV into a text portion PT by the voice recognition module 6 20 and of the translation of a text portion PT into a text portion translated PTT by the module 7, a vocal portion PV is processed several times by the functional means 4, 5, 6, 7, 10 and 11. For example, a passage of a vocal portion 25 twice to K through the means 4, 5 , 6, 7, and 11 refines the relevance of the contexts of the text segments deduced from this vocal portion. The number K of processing cycles of a voice portion depends on the time constraints, on the quality of each processing in means 4, 5, 6, 7, 10 and 11 and on the capacity of the buffer memory in the unit 11.

  The language converter CL comprises at least one translation module 7 and at least one speech synthesizer 8. The translation module 7 is activated when the unit 4 finds that the second language designated by the language identifier IL read in correspondence with the user identifier IU and the equipment EQm in the database BD is different from the language of the current voice portion PV determined by the unit 4. The translation module 7 translates the text portions PT into portions text texts translated PTT into said designated language and applied on the one hand to the unit 11 to enrich the contextual database 10 which is preferably multilingual and on the other hand to the voice synthesizer 8. Preferably, the voice recognition module 6 and the translation module 7 use a common context analysis to reduce processing times.

  Like the textual portions PT in the first language, the textual portions PTT translated into the second language can be applied to the unit 11 to be segmented into periodic time text segments in order to determine according to the respective speech parameters PVS and according to translated text segments the contexts of the translated text segments which enrich the contextual database 10. The translated text portion PTT is then submitted to the speech synthesizer 8. The speech synthesizer 8 transforms the translated text portion PTT into a synthesized portion PS when the type T of the translated text portion PTT must be converted and therefore does not constitute a portion recognized in the database 2 by the analyzer 1, as a replaced portion translated PRT or an untranslated portion PNT. Consequently, the analyzer 1 inhibits the operation of the speech synthesizer 8, and as a variant the chain of modules 4, 6 and 7, in response to 2 4 each PNT, PRT portion detected in the incoming audio signal SA and known to the database 2.

  For speech synthesis, the speech synthesizer 8 comprises in particular a buffer memory for memorizing the text portion PTT to be synthesized. The speech synthesis in the synthesizer is based on the PVS voice parameters resulting from the continuous analysis of the PV voice portion by the voice analyzer 4 and supplied continuously, and preferably on the linguistic parameters PL read in the database. BD data when they have been chosen by the user.

  When the voice analyzer 5 determines voice parameters characterizing sections respectively allocated to speakers and included in a PV voice portion, the synthesizer 8 continuously synthesizes the respective translated text portion PTT according to the PVS voice parameters. The voice synthesizer 8 decreases or optionally increases the bit rate of the synthesized portion PS by 10% in order to reduce its duration difference with the translated text portion.

  Because of speech synthesis and translation, the synthesized portion PS has a different duration from the corresponding original speech portion PV in the audio signal SA. This difference in duration depends mainly on the language of translation chosen by the user.

  The residual portion PR has the same duration as the corresponding vocal portion PV. The role of the mixer 9 is to transform the residual portion PR and the synthesized portion PS having different durations into two portions which have a common duration less than the greater of the durations of the residual portion PR and of the respective synthesized portion PS , preferably between the durations of the PR and PS portions. For this synchronization, the mixer 9 determines the mixing parameters characterizing an audio signal, such as the fundamental rate or frequency and / or the harmonics of the audio signal. By acting on the mixing parameters of each residual portion PR and of the synthesized portion PS, the mixer increases the duration of one of the portions of the residual and synthesized signal, in general the residual portion, and decreases the duration of the another portion, generally the synthesized portion, so that the two portions have the same duration and are in phase. The mixer then mixes the two portions of the same duration and in phase into a mixed portion PM of the converted audio signal SAC.

  To obtain a residual portion PR and a corresponding synthesized portion PS of the same duration by acting on the mixing parameters, for example an average of the durations of the two portions is determined by the mixer. The mixer acts on the mixing parameters of each of the residual and synthesized portions so that the durations of the residual and synthesized portions are equal to the average duration, one of the portions being compressed in time and the other portion being relaxed in time. The action by the mixer on the mixing parameters is limited. In fact, the synthesized and residual portions must keep the "audible" character of the corresponding initial portions for the human ear. To this end, the variation in the duration of these portions does not exceed a predetermined percentage, for example 10%.

  As a variant, the mixer acts only on the mixing parameters of the synthesized portion PS 35 so that the duration of the latter is equal to the duration of the residual signal portion. Many other variants fall within the scope of the invention in order to obtain corresponding residual and synthesized portions having the same duration, when these two portions are initially of different durations.

  In a first case where the voice portion PV is greater than a predetermined duration, approximately 2 minutes, the translation time by the system is too long. In a second case where the predetermined percentage is not sufficient to equalize the durations of the synthesized portion PS and of the residual portion PR, the mixer cannot process the synthesized portion PS and the corresponding residual portion PR in the same state. In both cases the mixer segments the synthesized portion and the corresponding residual portion into the same number of segments so that the mixer processes each synthesized segment and the corresponding residual segment and no longer the portions.

  In order to prevent the monotonous character of the voice resulting from the speech synthesis, the mixer 9 has several means such as the mixing of an additional noisy signal with the synthesized portion PS. This additional noisy signal is a portion of song contained in the audio database 2 and preferably chosen as a function of the contexts of the synthesized portion PS. The duration of the additional noisy signal is substantially equal to that of the corresponding synthesized portion. The mixer performs the operation of transforming the synthesized portion PS and the additional noisy signal into a mixed portion as before for the synthesized portion PS and the residual portion PR. A second means in the mixer consists in modifying the vocal parameters of the synthesized portion when these have not varied for a predetermined time, of the order of 90 seconds.

  The temporal adaptation unit 12 receives on the one hand the mixed portions PM and on the other hand the non-replaced portions PNT and replaced PRT which are written in parallel with the audio signal SA in a buffer memory in the unit 12 The conversion time of the portions to be converted PAC by the system being longer than that of the portions not replaced or replaced, the adaptation unit delays the portions not replaced PNT and replaced PRT so that the portions PNT and PRT are joined to the PM portions in the same order as the initial PNT, PRT and PAC portions 1 5 at the audio analyzer 1 in order to continuously compose the converted audio signal SAC at the output of the unit 12.

  When a first sum of the durations of the portions of the converted signal SAC is greater than a second sum of the durations of the corresponding portions of the incoming audio signal SA, the adaptation unit 12 reduces the duration of the converted output signal SAC by controlling the substantially equal suppression of portions of the audio signal SA having a duration substantially equal to the difference between the first and second sums by a suppression signal SS applied to the audio analyzer 1 which transmits the audio signal SA in the form of a signal audio restricted SAR. The portions to be deleted are preferably selected by the analyzer 1 as a function of their type T, such as for example commercials posterior to the signal SS and having a duration substantially equal to the difference between the first and second sums. Conversely, when the first sum of the durations of the portions of the converted signal SAC is less than the second sum of the durations of the corresponding portions of the audio signal SA, the adaptation unit 12 increases the duration of the converted signal SAC by adding 5 audio portions to the converted signal, having a duration substantially equal to the difference between the second and first sums to fill the time deficit of the converted signal with respect to the incoming audio signal and thus remove silences in the converted signal detected by the unit 12. These audio portions are extracted from pieces contained in the audio database 2 as a function, for example of their duration, and of the context of the portions of the audio signal SA.

  The converted audio signal SAC is then sent continuously to the terminal installation of the IT user.

  In another embodiment, the audio signal SA 20 is processed in different languages by the system even before it is broadcast. The system then offers real-time translation seen by the user.

Claims (9)

  1 - System for converting a continuous incoming audio signal (SA) including a voice signal in a first language into an audio signal converted in a second language, said system comprising means (1) for segmenting the received audio signal (SA) in audio portions (PAC), a means (3) for filtering the audio portions (PAC) each into a vocal portion (PV) and a residual portion (PR) of the same duration, a voice recognition means (6) converting the voice portions (PV) into text portions (PT) in the first language, a means for translating (7) the text portions (PT) 15 into translated text portions (PTT) in the second language, and a means (8 ) to synthesize the translated text portions (PTT) by voice into synthesized portions (PS) having different durations respectively from the audio portions, characterized in that it comprises a mixing means (9) for transforming each residual portion (PR) and the portion respective synthesized (PS) into portions which have a duration shorter than the greater of the durations of the residual portion (PR) and of the respective synthesized portion (PS) and which are in phase in order to mix them into a mixed portion ( PM) composing the converted audio signal (SAC).   2 - System according to claim 1, 30 comprising means (12) for detecting a temporal excess of a sum of the durations of the portions of the converted signal (SAC) relative to a sum of the durations of the corresponding portions of the incoming audio signal ( SA) so that the means for segmenting (1) deletes a portion of the incoming audio signal whose duration is substantially equal to said time surplus.   3 - System according to claim 1 or 2, 5 comprising means (12) for detecting a time deficit of a sum of the durations of the portions of the converted signal (SAC) relative to a sum of the durations of the corresponding portions of the audio signal incoming (SA) so that the means for detecting (12) 10 adds a portion to the converted signal (SAC) whose duration is substantially equal to said time deficit.   4 - System according to claim 1 to 3, wherein the means for segmenting (1) delimits in the incoming audio signal (SA) portions (PNT, PRT) which are contained in the database (2) and d other audio portions (PAC) which are applied to the filter means (3).   5 - System according to claim 4, wherein portions contained in the database (2) and delimited in the incoming audio signal (SA) are replaced respectively by replacement portions (PRT).   6 - System according to claim 4 or 5, wherein the means for segmenting (1) inhibits the operation of the means for synthesizing (8) in response to each portion (PNT, PRT) in the known incoming audio signal (SA) from the database (2).   7 - System according to any one of claims 1 to 6, comprising means (5) for analyzing the vocal portions (PV) to determine speech parameters (PVS), means (11) for segmenting the text portions (PT, PTT) in the first language, or in the first and second languages, in periodic temporal text segments, and means (11) for determining contexts (CSn) of the text segments according to the respective speech parameters (PVS ) and according to the text segments.   8 - System according to claim 7, in which the means for analyzing (5) determines vocal parameters characterizing sections respectively allocated to speakers and included in a vocal portion (PV) so that the means for synthesizing (8) synthesizes continuously the respective translated text portion 15 (PTT) according to the voice parameters (PVS).   9 - System according to any one of claims 1 to 8, included in a server (SE). 20 - System according to any one of claims 1 to 8, included at least partially in a user terminal installation (IT).