Method and system for a parametric speech synthesis
Embodiments of the present systems and methods may provide techniques for synthesizing speech in any voice in any language in any accent. For example, in an embodiment, a text-to-speech conversion system may comprise a text converter adapted to convert input text to at least one phoneme selected from a plurality of phonemes stored in memory, a machine-learning model storing voice patterns for a plurality of individuals and adapted to receive the at least one phoneme and an identity of a speaker and to generate acoustic features for each phoneme, and a decoder adapted to receive the generated acoustic features and to generate a speech signal simulating a voice of the identified speaker in a language.
1 . A speech conversion system comprising;
a memory for storing a plurality of phonemes;
a text converter adapted to convert input text to at least one phoneme selected from the plurality of phonemes stored in the memory, wherein the text converter is further adapted to detect a language of the input text;
a machine-learning model storing voice patterns for a plurality of individuals and adapted to receive the at least one phoneme from the text converter and an identity of a speaker and to generate and enhance acoustic features for each of the at least one phoneme, wherein the machine-learning model comprises a deep learning neural network that uses (i) phoneme embedding, (ii) speaker embedding, and (iii) language embedding, and wherein the voice patterns comprise a plurality of production components and a plurality of acoustic components arranged in a matrix, each of the plurality of production components contributing a selected amount to each of the acoustic components, such that the selected amount contributed is present in the matrix for each pair of production component and acoustic component, and wherein the enhanced acoustic features comprise at least one of spectral enhancement or focal enhancement; and
a decoder adapted to receive the generated acoustic features from the machine-learning model and to generate a speech signal simulating a voice of the identified speaker in the language, wherein the generated acoustic features include accent acoustic features such that the generated speech signal further simulates the voice of the identified speaker in the language and in an accent.
2 . The system of claim 1 , wherein the at least one phenome comprises a phoneme of the International Phonetic Alphabet and silence and breath.
3 . The system of claim 1 , wherein the machine-learning model comprises a neural network model.
4 . The system of claim 1 , wherein spectral enhancement comprises increasing a peak of a spectral envelope or decreasing a trough of the spectral envelope and focal enhancement comprises emphasizing the difference between a first frame and a second frame.
5 . The system of claim 1 , wherein the generated acoustic features include accent acoustic features and the generated speech signal further simulates a voice of the identified speaker in a language and in an accent.
6 . The system of claim 5 , wherein the accent corresponds to a native accent of the identified speaker.
7 . The system of claim 1 , wherein the plurality of production components comprise phones, coarticulation, prosody from linguistic features, and prosody from extra-linguistic features, and the plurality of acoustic components comprise spectrum, power, duration, and pitch.
8 . A speech conversion method implemented in a computer system comprising a processor, memory accessible by the processor, and computer program instructions stored in the memory and executable by the processor, the method comprising:
storing a plurality of phonemes in the memory;
converting input text to at least one phoneme selected from the plurality of phonemes stored in the memory;
detecting a language of the input text;
storing, in a machine-learning model at the computer system, voice patterns for a plurality of individuals, receiving, at the machine-learning model at the computer system, the at least one phoneme and an identity of a speaker, and generating and enhancing, with the machine-learning model at the computer system, acoustic features for each of the at least one phoneme, wherein the machine-learning model comprises a deep learning neural network that uses (i) phoneme embedding, (ii) speaker embedding, and (iii) language embedding, and wherein the voice patterns comprise a plurality of production components and a plurality of acoustic components arranged in a matrix, each of the plurality of production components contributing a selected amount to each of the acoustic components, such that the selected amount contributed is present in the matrix for each pair of production component and acoustic component, and wherein enhancing acoustic features comprises at least one of spectral enhancement or focal enhancement; and
receiving, at the computer system, the generated acoustic features from the machine learning model, and generating, at the computer system, a speech signal simulating a voice of the identified speaker in the language, wherein the generated acoustic features include accent acoustic features such that the generated speech signal further simulates the voice of the identified speaker in the language and in an accent.
9 . The method of claim 8 , wherein the at least one phenome comprises a phoneme of the International Phonetic Alphabet and silence and breath.
10 . The method of claim 8 , wherein the machine-learning model comprises a neural network model.
11 . The method of claim 8 , wherein spectral enhancement comprises increasing a peak of a spectral envelope or decreasing a trough of the spectral envelope and focal enhancement comprises emphasizing the difference between a first frame and a second frame.
12 . The method of claim 8 , wherein the generated acoustic features include accent acoustic features and the generated speech signal further simulates a voice of the identified speaker in a language and in an accent.
13 . The method of claim 12 , wherein the accent corresponds to a native accent of the identified speaker.
14 . The method of claim 8 , wherein the plurality of production components comprise phones, coarticulation, prosody from linguistic features, and prosody from extra-linguistic features, and the plurality of acoustic components comprise spectrum, power, duration, and pitch.
15 . A speech conversion system comprising;
a memory for storing a plurality of phonemes;
a text converter adapted to convert input text to at least one phoneme selected from the plurality of phonemes stored in the memory, wherein the text converter is further adapted to detect a language of the input text;
a machine-learning model storing voice patterns for a plurality of individuals and adapted to receive the at least one phoneme from the text converter and an identity of a speaker and to generate and enhance acoustic features for each of the at least one phoneme, wherein the machine-learning model comprises a deep learning neural network that uses (i) phoneme embedding, (ii) speaker embedding, and (iii) language embedding,
wherein the voice patterns comprise a plurality of production components and a plurality of acoustic components arranged in a matrix, each of the plurality of production components contributing a selected amount to each of the acoustic components, such that the selected amount contributed is present in the matrix for each pair of production component and acoustic component, and having voice pattern options, and wherein the enhanced acoustic features comprise at least one of spectral enhancement or focal enhancement; and
a decoder adapted to receive the generated acoustic features from the machine-learning model and to generate a speech signal simulating a voice of the identified speaker in the language, wherein the generated acoustic features include accent acoustic features such that the generated speech signal further simulates the voice of the identified speaker in the language and in an accent.