The MULTIVOC Text-To-Speech System

In this paper we introduce MULTIVOC, a real-world text-to-speech product geared to the French language. Starting from a ordinary French text, MULTIVOC generates in real-time a high quality speech using a synthesis-by-diphone method. The processing is divided into 3 main transformations (phonetization, automatic prosody and rhythm marking, and generation of LPC frames). This paper provides a full description of MULTIVOC including not only the technical view but also some applications of the product within the real world. 1. PRESENTATION OF MULTIVOC The text-to-speech MULTIVOC system is the result of a technology transfer from a research institute (CNET Lannion, France), which developed the basis of the system, to an industrial company (Cap Sogeti Innovation, France) which made the system a commercial product. Generating Linear Prediction Coding frames from ordinary text written in French, the goal of MULTIVOC is to give any standard applications the ability to produce (in real time) low-cost and high-quality speech output. MULTIVOC is shipped as a complete software system which aims to provide a sophisticated driver enabling applications to directly send French spoken text. The software package consists of the kernel of the driver itself and a set of dictionaries used by it. Several tools in the package allow an advanced user to tailor his own MULTIVOC driver to specific usage. Beside this static configuration facility, MULTIVOC also provides several run-time features. By submitting specific requests an application can change the following parameters: • The sampling frequency for generated frames. Three different frequencies are available: 8 kHz, 10 kHk and 16 kHz. This parameter will characterize the quality of the output voice, a frequency of 16 kHz providing the best results. • The tone of the output voice can be adjusted in the range 50-350 Hz. • The speech speed may be set from 1 to 10 syllables per second. • Two styles of prosody are provided. The "reading-style" corresponds to the usual way of reading a text, while the "advertising-style" is dedicated to short commercial messages like jingles. • One can also choose between a female or a male voice. The method used for the synthesis produces Linear Prediction Coding (LPC) frames generated from a diphone dictionary. Such a dictionary is specific to the sampling frequency used (8, 10 or 16kHz) and also to the style of voice (Female or Male). For this purpose, MULTIVOC provides 6 differents diphone dictionaries. The overall processing is organized as a pipelined set of transformations applied to the input text. At the higher level, one can distinguish the following functions: The pre-processing (or lexical processing) is a text-to-text transformation aiming to expande some non-worded terms like numbers (1987 --> "Mille Neuf Cent Quatre-Vingt-Sept"), administrative numbers (A4/B5 --> "A Quatre B Cinq") or acronyms (CSINN. --> "Cap Sogeti Innovation"). The phonetization process transforms the pre-processed text into phonemes according to predefined rules stored in a user-modifiable base. The prosody marking process scans the phonetized text and generates appropriate marks to reflect the prosody of the text using built-in rules based on the different punctuation signs and the grammatical type of words. The rhythm marking process computes the duration associated to each phoneme. Last, the frame generation process produces the LPC frames which correspond to the input text according to the different parameters specified and can be sent directly to the output device.